Optical coherence tomography enables optical biopsy of endometrial tissue for early cancer detection.

Optical coherence tomography enables optical biopsy of endometrial tissue for early cancer detection - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Npj Imaging . 2026 Jun 3;4:39. doi: 10.1038/s44303-026-00160-z Optical coherence tomography enables optical biopsy of endometrial tissue for early cancer detection Sanskar Thakur 1, 2, # , Yixiao Lin 2, # , Jinhua Xu 1, 2 , Haolin Nie 2 , Sarah Badwan 3 , Lukai Wang 1, 2 , Brooke E Sanders 3 , Premal H Thaker 3 , Andrea R Hagemann 3 , Carolyn K McCourt 3 , Dineo M Khabele 3 , Matthew A Powell 3 , David G Mutch 3 , Lindsay M Kuroki 3 , Ian S Hagemann 3, 4 , Quing Zhu 1, 2, 5, ✉ Author information Article notes Copyright and License information 1 Imaging Science Program, Washington University in St. Louis, St. Louis, MO 63130 USA 2 Department of Biomedical Engineering, Washington University in St. Louis, St. Louis, MO 63130 USA 3 Department of Obstetrics and Gynecology, -Division of Gynecologic Oncology, Washington University School of Medicine, St. Louis, MO 63110 USA 4 Department of Pathology and Immunology, Washington University School of Medicine, St. Louis, MO 63110 USA 5 Department of Radiology, Washington University School of Medicine, St. Louis, MO 63110 USA ✉ Corresponding author. # Contributed equally. Received 2025 Dec 28; Accepted 2026 Mar 31; Collection date 2026. © The Author(s) 2026, modified publication 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13233832  PMID: 42236613 Abstract Endometrial cancer is the most common malignancy of the female reproductive tract, yet diagnosis relies on invasive tissue sampling with risk of false negatives. We report an initial investigation of catheter-based 3D optical coherence tomography (OCT) integrated with functional, structural, and radiomic feature analysis for noninvasive endometrial assessment in ex vivo specimens. A custom 3.1-mm OCT catheter acquired volumetric images from 57 intact uterine specimens spanning normal endometrium, benign conditions, high-risk endometrial intraepithelial neoplasia, and endometrial cancer. Functional, structural, and radiomic features were extracted from OCT intensity and scattering images. Statistically significant features were selected to construct a cosine similarity matrix and a cosine-similarity network graph. Cosine-similarity visualization demonstrated exploratory sensitivity and specificity of 94% and 87%, while a cross-validated Logistic Regression Classifier yielded sensitivity and specificity of 91% and 83%. These findings support catheter-based 3D OCT as a promising noninvasive optical biopsy approach to improve detection of endometrial cancer. Subject terms: Cancer, Diseases, Medical research, Oncology Introduction Endometrial cancer (EC) is the most common gynecologic malignancy in the United States, with 69,120 diagnoses and 13,860 deaths projected in 2025, and incidence rising 1–3% annually 1 . Endometrioid adenocarcinoma is the predominant subtype, and early detection with timely treatment is critical. The 5-year survival rate reaches 80–90% for stage I EC but declines to 50–65% for stage II and 15–17% for stage IV 2 . Endometrial hyperplasia, marked by abnormal endometrial growth, carries a cancer risk from 2% in non-atypical hyperplasia to 29% 3 , or up to 42% 4 , 5 in atypical hyperplasia/endometrial intraepithelial neoplasia (EIN). Early diagnosis and medical management of EIN may preserve fertility and avoid hysterectomy 4 , 6 , yet no noninvasive screening tests currently exist. Transvaginal ultrasound (TVUS), while useful for detecting uterine abnormalities, is ineffective for early EC 7 . Endometrial biopsy, though safe, has a 10% false-negative rate due to sampling and interpretive variability 8 , 9 . Minimally invasive approaches using blood, uterine lavage, or cervicovaginal fluid have been explored, but evidence remains limited 10 . Hysteroscopy allows direct visualization and targeted biopsy but lacks subsurface information on endometrial and glandular architecture. A real-time, noninvasive, high-resolution modality for subsurface imaging could improve diagnostic accuracy, reduce unnecessary biopsies, and support fertility-sparing management. By avoiding repeated biopsies and downstream effects like inflammation and scarring, such a tool would be particularly valuable for women receiving progestin therapy and undergoing serial biopsies 11 . Optical coherence tomography (OCT) is a high-resolution imaging technique that utilizes low-coherence interferometry to capture cross-sectional images of subsurface biological tissues with micrometer-scale resolution up to 1–2 mm depth. By measuring the echo time delay and intensity of backscattered light, OCT allows for real-time imaging of tissue micro-structure in vivo. OCT’s resolution is 200 times higher than that of TVUS and approaches that of histopathological examination 12 . OCT has been at the forefront of research and application in ophthalmology 13 , 14 , cardiology 15 , dermatology 16 , gastroenterology 17 – 19 . OCT has also been explored for imaging female reproductive organs including cervix 20 – 22 , fallopian tubes 23 , 24 , and ovaries 25 , 26 . In this study, we report, to the best of our knowledge, the first catheter-based 3D OCT imaging integrating optical functional, structural, and radiomic features to assess the endometrial cavity. Images were acquired from 57 freshly excised hysterectomy specimens representing diverse pathological conditions. We introduce a cosine similarity matrix and a novel cosine-similarity network graph computed from feature space to visualize clustering of normal/benign and EIN/EC specimens, and employ a simple Logistic Regression Classifier to categorize specimens as normal/benign or EIN/EC. Results The analysis included 57 patient specimens: 16 premenopausal and 41 postmenopausal. (Table 1 ). There were 23 grade 1 EC cases, 6 grade 2 EC cases, 5 EIN cases, 1 hyperplasia without atypia case, 7 cases with cystic atrophy and benign polyp, and 15 normal cases. Table 1. Summary of specimen characteristics and pathological diagnosis after surgery Patient groups ( n = 57) Mean Age(y): 60 Range: 33–86 Histological diagnosis based on Hematoxylin and Eosin (H&E) Malignant ( n = 29) EC only ( n = 21) EC with EIN ( n = 8) 63 (35–86) FIGO grade 1 ( n = 23) FIGO grade 2 ( n = 6) EIN ( n = 5) 60 (47–83) EIN ( n = 2) EIN with treatment effects ( n = 3) Benign ( n = 8) Pre-menopausal ( n = 1) Post-menopausal ( n = 7) 64 (47–77) Non-atypical hyperplasia ( n = 1) Polyps and cystic atrophy ( n = 7) Normal ( n = 15) Pre-menopausal ( n = 9) Post-menopausal ( n = 6) 52 (33–81) Proliferative endometrium ( n = 8) Inactive endometrium ( n = 5) Secretory endometrium ( n = 1) Asynchronous pattern ( n = 1) Open in a new tab Examples of 2D OCT images (B-scans) alongside representative histological images from the same specimens are presented in Fig. 1 . In a 36-year-old premenopausal woman with normal proliferative endometrium (Fig. 1a ), the OCT images showed dense uterine glands with tubular patterns, which correlated with the coil-like appearance observed by histology. In a 47-year-old premenopausal woman with benign non-atypical hyperplasia, the OCT B-scan revealed dense glands appearing as low-intensity ellipsoids (Fig. 1b ). In a 49-year-old woman with EIN, the OCT images showed dilated glands with complex shapes and varying sizes, characterized by very low OCT intensity (Fig. 1c ). In a 64-year-old woman with grade 1 EC, the OCT images revealed merged, complex glands with bright lines, indicative of dense cellular structures (Fig. 1d ), which correlated well with the packed glandular architecture and neoplastic cells seen in the representative histology. Fig. 1. Examples of OCT B-scans. Open in a new tab a Normal endometrium b Endometrium with non-atypical hyperplasia (benign), c Endometrial Intraepithelial Neoplasia (EIN), and d endometroid adenocarcinoma with subsections showing representative OCT B-scan ( a1 , b1 , c1 , d1 ); representative H&E images ( a2 , b2 , c2 , d2 ); and cropped rectangular scans ( a3 , b3 , c3 , d3 ). The arrows highlight similar structures observed in H&E and OCT images. All scale bars are 1 mm. For the same specimens, 3D OCT images (C-scans) acquired at 150 µm depth from the tubing surface are presented in Fig. 2 , along with corresponding 3D OCT optical attenuation coefficient (µ t ) and scatterer shape factor (〈 D 〉) maps for the boxed regions. In the normal endometrium, homogeneous structures were observed, with small low-intensity dots and tubular patterns representing glands opening to the surface, on a background of endometrial stroma (Fig. 2a1 ). In benign hyperplasia, some regions resembled the normal endometrium seen in Fig. 2a1 , with low-intensity dots and tubular patterns, while other regions showed enlarged tubular structures (Fig. 2b1 ). In EIN, the 3D OCT image revealed more complex, low-intensity regions, indicative of irregular, dilated glands of various sizes (Fig. 2c1 ). In EC, the intensity map near the fundus showed distorted architecture with clustered irregularities and heterogeneous structures (Fig. 2d1 ). The low-intensity regions on the sides and in the middle-represented areas of low contact due to the circular shape of the catheter probe. Fig. 2. Examples of OCT C-scans. Open in a new tab a Normal endometrium, b non-atypical hyperplasia (benign), c Endometrial intraepithelial neoplasia (EIN), and d Endometrial cancer (EC). a1 , b1 , c1 , d1 (scale bar = 5 mm) OCT C-scan; a2 , b2 , c2 , d2 optical attenuation coefficients (µ t ) in cm -1 of the highlighted sections; a3 , b3 , c3 , d3 scatterer shape factors ( 〈 D 〉 ) of the highlighted sections. Images are oriented with the anterior on the right, posterior on the left, fundus at the top, and endocervix at the bottom. Compared to normal endometrium and benign hyperplasia, EIN and EC showed higher irregularity and heterogeneity in their µ t and <D> maps. In EIN, local rise in µ t and variations in 〈 D 〉 match with regions with enlarged irregular glands or complex mucosal folding. EC has distorted and fused glands. As a result, the typical glandular pattern was less obvious or absent in optical intensity map, but the heterogeneity became more pronounced. Feature analysis on samples Selected optical functional, structural and radiomic features that most effectively differentiate EIN and EC from normal and benign endometrium are summarized in Fig. 3 . In this statistical analysis (Fig. 3 and Table 1S ) anterior and posterior sides of each specimen were analyzed as independent samples. We refer “sample” as either anterior or posterior side, and “specimen or uterus” as both sides in the following manuscript. Analyzing these two sides effectively doubles the number of observations and identifies statistically significant features associated with regional endometrial lesions. Both regions have an equal likelihood of harboring disease. Fig. 3. Selected functional, structure, and radiomic features. Open in a new tab Selected optical functional features derived from the optical attenuation coefficient (µt) and the scatterer shape factor (D) maps are shown in ( a , b ) and ( c , d ), respectively. Selected optical structural features are shown in ( e , f ), and radiomic features computed from the OCT intensity images are presented in ( g , h ). Each panel displays a swarm plot together with the corresponding box plot. Each point represents a feature value from the anterior or posterior section of an individual endometrium. Welch’s t-test (*** p < 0.001) are shown above each comparison. A clear difference between normal/benign and endometrial intraepithelial neoplasia (EIN)/endometrial cancer (EC) groups can be observed in the plots. Optically, both EC and EIN exhibit increased spatial variation in optical scattering (Fig. 3a, b ) and decreased consistency in estimated scatterer size (Fig. 3c, d ), suggesting disrupted cellular organization and altered tissue composition. Structurally, EC and EIN demonstrate higher entropy in tissue spectral directionality (Fig. 3e ), consistent with the loss of normal architecture. Glandular ellipsoidal patterns with eccentricity values approaching 1 best represent the morphology of normal endometrial glands; and as shown in Fig. 3f , EC and EIN display significantly reduced similarity to this shape template, suggesting glandular distortion. Radiomic features further highlight these differences. EC and EIN exhibit lower homogeneity and higher entropy in pixel intensity (Fig. 3g, h ), reflecting decreased stromal uniformity. Specimen visualization and statistics Twenty-six statistically significant features were used to compute a 2D cosine-similarity matrix of 57 × 57 by taking inner product of L2-normalized high-dimensional feature vectors for each uterus. This matrix was used for specimen-level visualization, statistical analysis, and regression classifier cross-validation using the leave-one-out strategy. A novel 2D network graph based on cosine similarity for visualization of two clusters of normal/benign and EIN/EC cases is given in Fig. 4 . In the graph, each node represents a specimen and is randomly initialized on a unit circle to avoid layout bias. An edge is drawn when similarity meets or exceeds a chosen threshold. Fig. 4. A cosine similarity-based network graph visualization of the feature space. Your browser is not supporting the HTML5 <video> element. You may download that video as a file and play it with a player of you choice. Download video stream . Download video file (3.4MB, mp4) Open in a new tab A cosine similarity-based network graph visualization of the feature space with benign samples represented with circles (filled circles representing pre-menopausal samples), cancer represented in red crosses, and red triangles representing endometrial intraepithelial neoplasia (EIN). This video illustrates the evolution of a network graph built from 26 imaging features using cosine similarity. Each node represents a specimen; An edge is drawn when similarity meets or exceeds a chosen threshold; the animation sweeps the threshold monotonically from 1.0 down to 0.5. Nodes are randomly initialized on a unit circle to avoid layout bias. At frame 1 (threshold = 1.0), only identical specimens would connect, so no edges are present. As the threshold decreases toward 0.5, admissible neighborhoods expand, edge weights (cosine values) act as attractive forces in a force-directed layout, and each node’s position reflects the net of these forces. We stop at 0.5 to avoid adding weak, low-similarity links that would obscure structure. (In the PDF version of this article, the figure shows only the last frame of the video; please click anywhere on the figure or caption to play the video in a separate window). The network construction is unsupervised and relies solely on similarity of feature vectors; class labels are revealed only in the final frame to visualize separation between normal/benign and high risk (EIN) and cancer (EC) cases. By clustering similar nodes and separating weakly related ones, the graph produced clearer groupings and positioned cases with opposing features in the transition zone between normal/benign and EIN/EC clusters. A green dotted logistic regression boundary delineated the clusters. We observed correct classification in 32 of 34 EIN/EC cases and 20 of 23 normal/benign cases. The exploratory sensitivity and specificity are 94% and 87%, respectively. Cluster separation between the normal/benign and EIN/EC groups was supported by three metrics computed from cosine similarity matrix. The mean silhouette score of 0.28 indicates modest but non-random separation, consistent with partially overlapping feature distributions commonly observed in cancer imaging radiomics due to precancer and cancer heterogeneity and gradual disease progression 27 . The Calisnki-Harabasz (CH) index of 44.8 further suggests a meaningful clustering structure, reflecting a favorable ratio of between-group dispersion to within-group compactness. A permutation test, which yielded a P value of 0.001, indicates that the observed group structure is significant. Together, these results support the presence of distinguishable—though not perfectly separable—radiomic patterns between normal/benign and EIN/EC uteri (Table 2 ). Table 2. Cosine Similarity Measures Silhouette Score 0.28 Calinski–Harabasz Index (CH) 44.8 Permutation Test P = 0.001 Open in a new tab It is important for young patients to have accurate diagnosis on EC and EIN in order to preserve their uteri for future pregnancies. From the 10 premenopausal cases, only one specimen with proliferative endometrium was misclassified in the analysis based on the selected decision boundary. According to the pathology report, this specimen had a large portion of subendometrial area with hemorrhagic material which likely influenced the features and subsequent analysis. The rest of the nine cases were classified correctly. This highlights the potential benefit of OCT in EIN/EC screening and diagnosis of young patients. Two post-menopausal benign specimens were misclassified as cancer. These two cases had a cystic atrophy and multiple benign polyps, and the complex polyp and atrophic structure may have influenced the classification. Two of the 34 cancer specimens were misclassified as benign: one uterus had cancer at fundus and our 3D OCT analysis did not capture a significant portion of it while another was a treated cancer case with EIN, a partial responder that was misclassified likely due to the structural changes. Logistic regression classifier based on cosine similarity of specimen (Ss) A simple logistic regression-based classifier using leave-one-out strategy was cross validated on the 57 × 57 cosine similarity matrix computed from features. The classifier achieved AUC of 0.957. When a 0.5 threshold was used for the classifier, it achieved a sensitivity of 91.2% and a specificity of 82.6%, highlighting the separability of the dataset using cosine similarity derived from feature space. Three misclassified benign cases from the visualization were also mislabeled by the classifier. An additional post-menopausal specimen with very thin inactive endometrium was misclassified. For EC/EIN group, one misclassification with cancer at fundus was identical. Remaining two specimens had EC/EIN mixed with benign polyps and hyperplasia. This complex structure possibly contributed to the misdiagnosis. Discussion In this study, 3D OCT volumetric images were acquired from 57 uterine specimens spanning diagnoses from normal endometrium and EIN to FIGO grade 1–2 EC. 2D OCT images were interpreted using representative H&E histology from the corresponding anterior or posterior regions. Optical functional, structural, and radiomic features were extracted from 3D intensity images and scattering maps. Twenty-six statistically significant features were selected to compute a cosine similarity matrix and construct a novel network graph to visualize clustering of normal/benign and EIN/EC specimens. Cosine-similarity visualization showed moderate separability between clusters with exploratory sensitivity of 94% and a specificity of 87%. A Logistic Regression Classifier yielded AUC of 0.957 with a sensitivity of 91% and a specificity of 83% (Fig. 5 ). These findings demonstrate the potential of catheter-based 3D OCT as a new noninvasive “optical biopsy” to improve detection of endometrial cancer. Fig. 5. ROC of logistic regression classifier. Open in a new tab A receiver operating characteristic (ROC) curve from the Logistic Regression Classifier with leave-one-out cross validation (LOOCV). Existing literature on the use of OCT to assess the endometrium is limited. In a pilot study conducted by Pitris and Goodman et al. 28 OCT was investigated for the evaluation of endometrial neoplasms. The study demonstrated potential in differentiating various endometrial pathologies using a benchtop system, although imaging was very slow (10–30 s per image). New-generation OCT systems with fiber-optic catheter probes now make imaging of hollow organs and their internal lining, such as the endometrium, feasible at much faster rates while maintaining resolution comparable to benchtop systems. Recently, Ding et al. conducted an ex vivo evaluation of human endometrium using a commercial OCT system and showed that EC exhibits a high-density heterogeneous signal, corresponding to high cellular density, with good correlation between histology and OCT images 29 . However, their study was limited to two-dimensional (B-scan) intensity analysis and lacked three-dimensional reconstruction to evaluate the entire endometrial cavity. A similar study of 15 freshly resected specimens had comparable limitations in 3D imaging and analysis 30 . This study is the first to combine catheter-based 3D OCT imaging with functional, structural and radiomic feature analysis to assess the endometrial cavity, using data from 57 freshly excised hysterectomy specimens with varying pathological conditions. This study had some technical limitations. The side-viewing catheter with a sealed tubing tip used in the study blocked a few millimeters of the fundus area during each scan. As a result, one case with focal cancer in the fundus was misclassified as normal. Additionally, due to the circular shape of the catheter and the ellipsoid shape of the uterus, some areas on the lateral sides of the uterus (short axis) had poor contact with the catheter, resulting in very low intensity in the OCT images. By dividing and analyzing the 3D OCT volumetric images of the endometrial cavity into anterior and posterior sections for each uterus, we can avoid analyzing these poorly contacted regions. Another limitation of the study was the difficulty to co-register local histological sites with corresponding OCT images. Thus, the distinction between anterior and posterior sides was used to correlate histological observations with OCT. Future work will focus on optimizing catheter tip design to improve imaging of the fundus, developing advanced machine learning methods for tissue segmentation and lesion localization, and improving histological specimen processing. We also acknowledge the limited variability among the pre-menopausal normal samples in the study. Only one secretory endometrium was included, and no menstrual-phase endometrium was examined. Because menstrual cycle can alter the endometrial structure, further investigation is needed to evaluate how all menstrual phases affect OCT imaging in order to better characterize physiological variability and support clinical translation. The ability to image the entire endometrial cavity within minutes, generate 3D volumes with micron-level detail, and perform automated feature extraction and lesion diagnosis enables an ‘optical biopsy’ approach for timely noninvasive diagnosis and supports physicians in making critical clinical and surgical decisions. The next steps will be (1) increase the OCT data acquisition speed by incorporating a faster laser and increasing the catheter pull back speed; (2) optimize the OCT data processing pipeline; and (3) conduct patient studies aimed at translating OCT into a clinical optical biopsy tool for effective assessment of the endometrium and detection of EIN and endometrial cancer. Because the ex vivo endometrial specimens were imaged fresh immediately after surgery, tissue dehydration was unlikely, and we expect the OCT features observed ex vivo to translate well to in vivo analysis. Our side-viewing catheter provides good contact along the uterus’s long axis, motivating separate feature extraction from anterior and posterior regions, although limited contact can occur along the short axis. In future patient studies, real-time OCT feedback will allow image-quality assessment and catheter angle adjustment for a second pass to improve contact. In summary, this study demonstrates that 3D catheter-based OCT imaging, integrated with statistically significant functional, structural, and radiomic features, along with cosine similarity network visualization and logistic-regression classification, shows strong potential as a noninvasive optical biopsy tool for detecting high-risk endometrial intraepithelial neoplasia and endometrial cancer. This approach may complement standard biopsy practices and assist physicians in making important clinical and surgical decisions. Methods Patient recruitment Patients undergoing hysterectomy at Barnes-Jewish Hospital were enrolled under an ex vivo imaging protocol approved by the Institutional Review Board of Washington University School of Medicine (HRPO#:202409083). All patients signed informed consent. Inclusion criteria were normal endometrium, biopsy-proven hyperplasia without atypia, and EIN or low-grade 1 or 2 EC. Exclusion criteria included preoperative pelvic radiation, neoadjuvant chemotherapy for known EC, and conditions preventing OCT catheter insertion, such as extensive fibrosis. Written informed consent was obtained from all patients. A total of 69 specimens were imaged between December 2024 and June 2025. The first 12 cases were excluded while piloting workflow and optimizing imaging conditions. Final analysis included 57 uterine specimens, as shown in Fig. 6a . For this cohort of surgery patients, menopausal status and last menstrual period were self-reported. Some patients—likely those in the perimenopausal transition—did not report being post-menopausal. Based on clinical assessment, these individuals were classified as pre-menopausal. Fig. 6. Patient study flow chart, OCT system, and catheter tip. Open in a new tab a Study patient flow chart, b Optical coherence tomography imaging system; c Imaging catheter tip showcasing light path, and d Optical design of light delivery through the catheter tip. Catheter-based OCT system and Imaging Procedure Our previously designed catheter, measuring 3.1 mm in diameter and 1.5 m in length, together with an OCT imaging system operating at 1310 nm and 20 kHz 18 , was optimized to image the entire endometrial cavity (Fig. 6 ). A C + +/Python-based graphical user interface (GUI) was developed to operate the modified system. The catheter incorporated a single-mode optical fiber delivering light to a 1 mm gradient refractive index (GRIN) lens and 0.5 mm right-angle prism for tissue illumination. The lens was epoxied to the fiber tip and encapsulated in rigid stainless-steel tubing. The fiber was housed in a 0.082-inch helical hollow-strand stainless torque coil, shielded by 3.1 mm polytetrafluoroethylene tubing. A 2.9 mm transparent PEBAX sleeve (Teleflex OEM) was affixed at the tip to enhance transmission and enable rotation. The proximal end connected to the optical setup via a rotary joint (Princetel MJXA-155-28 T-002-FA), driven by a DC motor and mounted on a stepper-controlled linear stage. The system employed a swept-source laser (Santec HSL-2100, 1310 nm, 20 kHz), balanced detector (Thorlabs PDB450C), and high-speed acquisition card (AlazarTech ATS9462, 180 MS/s). The axial and lateral resolution of the system was ~6 µm and ~8.5 µm, respectively. The tip of the PEBAX tubing was heat sealed and no index matching fluing or saline was used during the study. Freshly resected uteri were imaged immediately after the surgery, before formalin fixation for histopathological evaluation, without any other preparation. Most samples were received with intact endometrial cavities and were imaged by inserting the OCT catheter through the cervix all the way to the fundus. The specimen was oriented to have anterior surface on the top and the physical marker on the probe tubing was aligned to be facing the anterior surface. The imaging catheter rotated inside the tubing at about 600 revolutions per minutes and automatically pulled back using a stepper motor at constant speed (0.25 mm/sec). Total 3–5 cm length of endometrial cavity was imaged based on the size of the uterus. A few samples ( n = 5) were received bivalved and were imaged by folding the anterior and posterior against the probe as naturally as possible. The 3D OCT data were processed and displayed in real time using a GUI with controls for acquisition, motor operation, and image display. Each 2D B-scan contained 2000 A-lines, continuously acquired using the 20 kHz laser trigger. A frame rate of 6 frames per second was achieved, and 3D scans were reconstructed by stitching sequential B-scans during catheter pullback. Entire endometrial cavities were imaged within 3–5 min. Data preprocessing and region of interest selection All the data preprocessing and analysis were performed on a 3D volume generated from flattened rectangular OCT B-scans. To minimize the impact of speckle noise on reliable tissue feature extraction, each B-scan was first denoised using a speckle-reducing bilateral filter 31 . Then each C-scan was resized so that the voxel size was 9.5 µm along all three axes. The volumetric OCT data set was divided into nonoverlapping volumes of interest (ROI) measuring 128 × 128 × 80 pixels in the radial, pullback, and depth directions, respectively, or 1.21 × 1.21 × 0.76 mm in physical dimensions. The ROI was measured on the 3D volume, with the x-axis running from left to right (posterior to anterior), the y -axis running from the fundus to the cervix, and the z-axis representing depth. The region of poor tissue contact was removed by identifying the tissue surface in each 2D rectangular B-scan using a standard method of generating a binary mask and identifying tissue surface after the tubing surface. 3D optical functional features Endometrial cancer cells are characterized by enlarged and irregular nuclei, a high nuclear-to-cytoplasmic ratio, prominent nucleoli, and increased mitotic activity. These histopathological alterations elevate the density and refractive index contrast of subcellular structures, increasing optical scattering. In OCT, this enhanced scattering manifests as a change in the optical attenuation coefficient, µ t , which quantifies the depth-dependent attenuation of the OCT signal. Therefore, µ t serves as a quantitative marker reflecting these underlying microstructural changes and provides critical information for cancer detection. For each A-line in a B-scan, µ t was estimated assuming the intensity of OCT signal I ( z ) follows the Beer-Lambert Law: I ( z ) ∝ e −2 µtz . To mitigate error due to tissue heterogeneity (e.g., glands and tissue folds), µ t was calculated as an average over small depth segments rather than a single value estimated from the entire A-line (Eq. 1 ) 32 . Specificaly, µ t was calculated for each A-scan using a sliding-window linear regression approach. The OCT intensity signal, I(z) , was divided into depth segments of 25 pixels, corresponding to a physical depth of approximately 200 µm. Within each segment, a linear fit of ln( I(z) ) versus depth z was performed to estimate the local attenuation slope, with µ t defined as –slope/2. The 25-pixel segment length was selected to balance resolution ( ~ 200 µm) with the stability of the linear regression against noise. Segment-wise µ t estimates were excluded before averaging if (1) they were negative, as can occur at bright boundaries (e.g., the stroma–gland interface) where intensity increases with depth, or (2) they exceeded 100 cm −1 . Although tissue scattering coefficients can surpass this value, the OCT-derived µ t approximates the scattering coefficient, and values above 100 cm −1 , roughly corresponding to a reduced scattering coefficient of 10 cm −1 in our data, typically occurred in regions of very low signal (e.g., inside glands or shadows), where the logarithmic fit became unstable and produced arbitrarily high estimates. The remaining physically plausible µ t values, restricted to the tissue region defined by a binary mask obtained via Otsu’s method 33 , were averaged to yield a single robust µ t estimate per A-scan. In addition, a scattering shape factor, D , a potential cancer marker characterizing optical scattering cross section, was estimated using the method proposed in ref. 34 . The approach assumes that the OCT signal power spectrum follows the power law in the wavelength domain, λ , such that I ( z , λ ) ∝ e − ( 4 − < D > ) ⋅ λ . Briefly, for each A-line, short-time Fourier transform (STFT) was first applied to partition the A-line fringe signal into nine (M = 9) equally spaced spectral windows between 1.22 um and 1.40 um, which were then reconstructed separately. STFT yielded a spectrogram ( I ( z , λ m ), m = 1,…,M) representing wavelength-dependent backscattering intensity. From the spectrogram, the spectral decay slope, α, was derived from linear regression of the logarithmic intensity across the range of wavelengths. Assuming the model ln I z , λ = − α λ + C , then the decay slope from a depth segment can be estimated by α z = − ∑ m = 1 M λ m ln I z , λ m − 1 M ∑ m = 1 M λ m · ∑ m = 1 M ln I z , λ m ∑ m = 1 M λ m 2 − 1 M ∑ m = 1 M λ m 2 1 The shape factor D z = 4 − α ( z ) was then obtained straightforwardly. Similar to the estimation of μ t , for each A-line, D was calculated as an average over valid small depth segments of 25 pixels, retaining only non-negative D z below 4 for averaging because both the spectral decay rate and the shape factor are constrained to be non-negative numbers. For each ROI, first-order statistical and gray level co-occurrence matrix (GLCM) based radiomic features were computed for µ t and <D> maps. Specifically, GLCM features were calculated based on the established method proposed by Haralick et al. 35 using the implementation from the skimage.feature library in Python. For each direction θ , a function-level GLCM G θ is constructed, where each element G θ [ i , j ] is calculated as the probability in the image that a pixel with function-level value i has a neighbor with value j that is at angle θ from the pixel. In our analysis, we discretized each map into 64 levels. We used θ ∈ { 0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ } with offset of 1 pixel, i.e., only immediate neighbors were considered. ROI-level features were then averaged across all ROIs inside each C-scan. These features are referred to as OCT functional features in the following text. 3D structural features To characterize endometrial tissue, two novel OCT structural features were derived: one from the spatial frequency domain and one from the OCT intensity image. Spatial Spectrum Features and Directionality: Structural descriptors were directly derived from 3D OCT intensity ROIs to detect global periodicity. First, a 3D spatial Fourier transform, F ( k x , k y , k z ) = F { I ( x , y , z ) } , of each 3D ROI was computed, and an ellipsoid was fitted to the amplitude spectrum. Parameters of the spectral ellipsoid, such as the major/minor axis length, eccentricity and orientation, have been shown to reflect periodic structures of tissue 36 . As healthy endometrial tissue has a homogeneous, regular glandular structure, the ellipsoid parameters are expected to have well-defined values. In contrast, endometrial cancer features crowded, enlarged, and complex glands, and the fitted ellipsoids would broaden and show higher variations. Further, the overall tissue directionality was determined by applying Radon transform to the amplitude spectrum as a function of spatial frequency orientation. To assess structural orientation in each anatomical plane, three orthogonal 2D projections of F ( k x , k y , k z ) were extracted by averaging the central planes along the perpendicular axis, yielding F ( k x , k y ) ® , F ( k y , k z ) ® and F ( k x , k z ) ® . Radon transform was then applied to each projected spectrum by integrating along lines at angle θ and distance ρ from the center of the spectrum, yielding R ( ρ , θ ) . Because Radon transform was applied to a spatial frequency spectrum, the center slice of R ( ρ , θ ) , or R ( 0 , θ ) , captures the angular distribution of spatial frequency, where peaks indicate dominant structural orientations. From this angular distribution, four statistics, i.e., standard deviation, skewness, kurtosis, and entropy, were extracted from each orthogonal plane. They were then averaged across all three orthogonal planes to yield the final quantitative measures of tissue directionality. Glandular Features of 3D OCT Intensity Images: To capture local glandular abnormalities, a library of 3D ellipsoidal templates was constructed from 3D OCT images, covering a physiologically plausible range of sizes, aspect ratios, and orientations. Each template was based on a normalized 3D Gaussian distribution E r ∼ e − 1 2 || Σ − 1 Rr|| 2 , where R is a 3D rotation matrix and Σ defines ellipsoidal scale along each intrinsic axes of the ellipsoid (not the global Cartesian axes). The library contained templates with shortest axes ranging from 2 to 16 voxels and aspect ratios, represented by the elongation factor, from 1 (spherical) to 9 (thin tubular). Orientations were sampled at 6 equally spaced azimuthal and polar angles, yielding a total of 6 ×6 = 36 unique ellipsoidal orientations. To obtain the 3D pattern matching response, each 3D ROI was convolved with every ellipsoid in the library. Because endometrial glands exhibit low signal intensities on OCT, only negative-valued convolution responses were attributed to true glandular structures, and the pattern matching fitness was evaluated based on their absolute values averaged over all 3D ROI from the same specimen. These novel structure features are referred to as OCT structure features in the following text. 3D radiomic features of OCT intensity image Finally, first-order statistics and GLCM-based radiomic features were computed directly on each 3D ROI of OCT intensity images. For each direction θ , a gray-level GLCM G θ is constructed, where each element G θ [ i , j ] is calculated as the probability in the B-scan image that a pixel with gray-level value i has a neighbor with value j that is at angle θ from the pixel. We discretized each B-scan into 64 levels and used θ ∈ { 0 ∘ , 45 ∘ , 90 ∘ , 135 ∘ } with offset of 1 pixel, i.e., only immediate neighbors were considered. The features calculated on B-scans were averaged over the 3D ROI, and then further averaged over the anterior/posterior after calculating for all 3D ROIs. The complete data processing pipeline is described in Fig. 7 . Fig. 7. Data processing pipeline describing feature extraction and analysis. Open in a new tab Optical functional features were extracted from optical attenuation coefficient ( µ t ) map and optical shape factor (D) map. Optical structural features were extracted from spectral domain structure regularity and directivity, and spatial domain glandular size and pattern. Radiomic features were extracted from intensity images. Sample statistics Endometrial lesions are frequently localized rather than uniformly affecting the entire endometrium. To improve classification accuracy and capture more detailed pathological changes, the anterior and posterior regions of each uterus scan were analyzed separately in statistical feature analysis as shown in Fig. 3 . By treating the anterior and posterior sides as distinct inputs, the analysis preserved spatial specificity and enhanced sensitivity to regional abnormalities. An imaging feature was considered significant for endometrial cancer diagnosis if it met both of the following criteria. First, the feature showed a statistically significant difference between normal/ benign and EIN/EC as indicated by Welch’s t -test ( p < 0.01). Second, the mutual information score between each feature and the percentage of endometrial cancer estimated from H&E slides by the study pathologist was computed using the mutual_info_regression function in scikit-learn package (Python 3.12). Features with scores above 0.05 were retained, indicating a non-negligible dependence on estimated malignancy. This feature selection procedure identified 26 statistically significant imaging features (supplementary Table S1 ). Cosine-similarity Matrix and Network Graph To qualitatively assess discriminative capacity of extracted high dimensional features, OCT feature space visualization and diagnosis that differentiates normal/benign from EIN/EC specimens were needed. While PCA and t-SNE are common methods for dimensionality reduction and visualization in high-dimensional data, PCA assumes linear relationships and focuses on preserving global variance 37 , and t-SNE is a nonlinear embedding designed to preserve local neighborhood relationships, but it is not designed to preserve global distances or topology, making high-dimensional relationships difficult to interpret from the low-dimensional map 38 . In contrast, a cosine-similarity network represents the relational structure among specimens directly in similarity space, providing a representation that can capture both local and broader class separation patterns that are particularly informative in small-sample radiomics datasets 39 . To better visualize specimen-level (patient-level) clustering based on OCT features, we computed a cosine-similarity matrix S and constructed a novel cosine similarity network graph using a force-directed layout 40 . A cosine-similarity matrix S s at specimen level was computed to assess the diagnosis of each specimen or uterus. Each 26-feature vector, derived from either the anterior or posterior region of a specimen, was L2-normalized prior to similarity computation. Cosine similarity between two feature vectors was calculated as their inner product. Each element of the matrix between specimens X and Y was calculated as the average of the four pairwise similarities: ( X a ⋅ Y a + X a ⋅ Y p + X p ⋅ Y a + X p ⋅ Y p ) / 4 2 where subscripts a and p denote anterior and posterior regions, respectively. In the cosine-similarity graph construction, each specimen was represented as a node, with edges connecting pairs of nodes weighted by the cosine similarity between their OCT feature vectors. In the development of cosine-similarity network, specimen labels were not used at any stage of graph construction or layout optimization; labels are only overlaid at the end for visualization and diagnosis. Cosine-similarity-based statistical analysis and specimen classification Statistical analysis was performed using S s for specimen-level classification. Silhouette Score: For each specimen X , the silhouette score s X = b X − a X max a X , b X , was computed, where a X and b X denote the mean distances to specimens in the same and the other class, respectively. The average silhouette score Si summarizes the overall separation between the two groups. Silhouette values near zero typically reflect overlapping clusters, whereas values above ~0.5 suggest well-separated groups 41 , 42 . In cancer radiomics studies, however, intrinsic precancer and caner heterogeneity and gradual transitions along disease continua often produce partially overlapping feature distributions, even when clinically meaningful cancer phenotypes are present 27 . Calinski–Harabasz Index (CH): The CH index evaluates clustering performance by quantifying the ratio of between-group variance to within-group variance. Higher CH values indicate well-separated and internally compact clusters. Permutation test: A permutation test was performed to robustly assess whether the observed differences in within- versus between-group cosine similarities could arise by chance. 1000 permutations were performed. A simple classifier with leave-one-out validation: A simple Logistic Regression Classifier was developed using a leave-one-out cross-validation strategy on the 57 × 57 cosine-similarity matrix derived from feature vectors. For each iteration, one specimen was held out for testing and four summary attributes were computed from the remaining 56 × 56 similarity matrix. For each training specimen, the mean and standard deviation of similarity to the normal/benign cluster and to the EC/EIN cluster were calculated, forming a 56 × 4 training matrix. The Logistic Regression model was trained 57 times on these matrices, and the held-out specimen was tested in each iteration to estimate overall performance. Subsequent pathology handling After OCT, uteri were returned to the pathology lab for routine processing per standard of care, including overnight formalin fixation, prosection by a pathologist or assistant, and preparation of paraffin-embedded sections stained with hematoxylin and eosin (H&E). For each patient, diagnosis was retrieved from the medical record and H&Es were reviewed by the senior pathologist of 14 years’ experience. When multiple diagnoses occurred (e.g., carcinoma and polyp), the most significant was retained for classification (e.g., carcinoma). Following the 2020 World Health Organization Classification of Tumors of the Female Genital Tract, atypical endometrial hyperplasia and endometrioid intraepithelial neoplasia (AEH/EIN) were treated synonymously 43 . Supplementary materials Supplementary Table 1S lists all 26 significant optical functional and structural features, and radiomic features along with the p -values between normal/benign and EIN/EC and the mutual information score of each feature with the pathology assessment of percentage of the EIN or EC. Supplementary information Supplementary information (8.9MB, docx) Acknowledgements This work was funded by the Developmental Research Program (DRP) of the NCI Route 66 Endometrial Cancer SPORE (5P50CA265793-03). Partial support for this work was provided by the NCI (R01CA237664) and the NIBIB (R01EB034398). Author contributions Conception and design: Q. Zhu, I.Hagemann, L. Kuroki. Development of methodology: S. Thakur, Y.Lin, J. Xu, H. Nie, I. Hagemann, L. Kuroki, Q. Zhu. Data acquisition: S. Thakur, J. Xu, H. Nie, S. Badwan, L. Wang, B.E. Sanders, P. Thaker, A.Hagemann, C. McCourt, D. Khabele, M. Powell, L. Kuroki; Analysis and interpretation of data: S. Thakur, Y.Lin, J. Xu, I. Hagemann, L. Kuroki, Q. Zhu. Statistic Review Y.Lin, Q. Zhu, Manuscript writing: S. Thakur, Y.Lin, I. Hagemann, L. Kuroki, Q. Zhu; Manuscript review I. Hagemann, L. Kuroki, Q. Zhu, D. Mutch, P. Thaker. All authors approved the submission. Data availability The patient datasets collected and analyzed for the reported study are not publicly available due to patient privacy concerns; however, the data can be made available upon request with the IRB’s permission. Code availability The underlying code to extract functional, structure and radiomic features from OCT data, and code of generating cosine similarity-based network graph are available in GitHub and can be accessed via this link https://github.com/OpticalUltrasoundImaging/OCTEndometrialCancer.git . Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Sanskar Thakur, Yixiao Lin. Change history 6/25/2026 In this article, Figure 4 needs to be processed as a video image, which has now been updated. Supplementary information The online version contains supplementary material available at 10.1038/s44303-026-00160-z. References 1. Key Statistics for Endometrial Cancer. https://www.cancer.org/cancer/types/endometrial-cancer/about/key-statistics.html . 2. Makker, V. et al. Endometrial cancer. Nat. Rev. Dis. Prim. 7 , 88 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Raglan, O. et al. Risk factors for endometrial cancer: an umbrella review of the literature. Int. J. Cancer 145 , 1719–1730 (2019). [ DOI ] [ PubMed ] [ Google Scholar ] 4. Kurman, R. J., Kaminski, P. F. & Norris, H. J. The behavior of endometrial hyperplasia. A long-term study of “untreated” hyperplasia in 170 patients. Cancer 56 , 403–412 (1985). [ DOI ] [ PubMed ] [ Google Scholar ] 5. Baak, J. P. et al. The molecular genetics and morphometry-based endometrial intraepithelial neoplasia classification system predicts disease progression in endometrial hyperplasia more accurately than the 1994 World Health Organization classification system. Cancer 103 (11), 2304–2312 (2005). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Kuhn, T. M., Dhanani, S. & Ahmad, S. An Overview of Endometrial Cancer with Novel Therapeutic Strategies. Curr. Oncol. 30 , 7904–7919 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. The Role of Transvaginal Ultrasonography in Evaluating the Endometrium of Women With Postmenopausal Bleeding https://www.acog.org/clinical/clinical-guidance/committee-opinion/articles/2018/05/the-role-of-transvaginal-ultrasonography-in-evaluating-the-endometrium-of-women-with-postmenopausal-bleeding . 8. Vitale, S. G. et al. Endometrial biopsy: indications, techniques and recommendations. An evidence-based guideline for clinical practice. J. Gynecol. Obstet. Hum. Reprod. 52 , 102588 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 9. Trimble, C. L. et al. Concurrent endometrial carcinoma in women with a biopsy diagnosis of atypical endometrial hyperplasia: a Gynecologic Oncology Group study. Cancer 106 , 812–819 (2006). [ DOI ] [ PubMed ] [ Google Scholar ] 10. Shen, Y., Yang, W., Liu, J. & Zhang, Y. Minimally invasive approaches for the early detection of endometrial cancer. Mol. Cancer 22 , 53 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Makris, N. et al. Three-dimensional hysterosonography versus hysteroscopy for the detection of intracavitary uterine abnormalities. Int. J. Gynaecol. Obstet. 97 , 6–9 (2007). [ DOI ] [ PubMed ] [ Google Scholar ] 12. Wessels, R. et al. Optical biopsy of epithelial cancers by optical coherence tomography (OCT). Lasers Med. Sci. 29 , 1297–1305 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Rusu, A. C. et al. Retinal structural and vascular changes in patients with coronary artery disease: a systematic review and meta-analysis. Life (Basel) 14 , 448 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Ahmed, H., Zhang, Q., Donnan, R. & Alomainy, A. Denoising of optical coherence tomography images in ophthalmology using deep learning: a systematic review. J. Imaging 10 , 86 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Yonetsu, T., Bouma, B. E., Kato, K., Fujimoto, J. G. & Jang, I.-K. Optical coherence tomography– 15 years in cardiology. Circ. J. 77 , 1933–1940 (2013). [ DOI ] [ PubMed ] [ Google Scholar ] 16. Olsen, J., Holmes, J. & Jemec, G. B. Advances in optical coherence tomography in dermatology-a review. J. Biomed. Opt. 23 , 1–10 (2018). [ DOI ] [ PubMed ] [ Google Scholar ] 17. Tsai, T. H. et al. Optical coherence tomography in gastroenterology: a review and future outlook. J. Biomed. Opt. 22 , 1–17 (2017). [ DOI ] [ PubMed ] [ Google Scholar ] 18. Nie, H. et al. In vivo evaluation of complex polyps with endoscopic optical coherence tomography and deep learning during routine colonoscopy: a feasibility study. Sci. Rep. 14 , 27930 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Bidani, K., Moond, V., Nagar, M., Broder, A. & Thosani, N. Optical imaging technologies and clinical applications in gastrointestinal endoscopy. Diagnostics (Basel) 15 , 2625 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Ren, C. et al. Multi-center clinical study using optical coherence tomography for evaluation of cervical lesions in-vivo. Sci. Rep. 11 , 7507 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Gan, Y. et al. Analyzing three-dimensional ultrastructure of human cervical tissue using optical coherence tomography. Biomed. Opt. Express 6 , 1090–1108 (2015). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Escobar, P. F. et al. Diagnostic efficacy of optical coherence tomography in the management of preinvasive and invasive cancer of uterine cervix and vulva. Int. J. Gynecol. Cancer 14 , 470–474 (2004). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Rocha, A. D. et al. Iterative prototyping based on lessons learned from the falloposcope in vivo pilot study experience. JBO 28 , 121206 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Luo, H. et al. Enhanced 3D visualization of human fallopian tube morphology using a miniature optical coherence tomography catheter. Biomed. Opt. Express 14 , 3225–3233 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Zeng, Y. et al. Histogram analysis of en face scattering coefficient map predicts malignancy in human ovarian tissue. J. Biophotonics 12 , e201900115 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Li, S., Luo, H., Kou, S., Hagemann, I. S. & Zhu, Q. Depth-resolved attenuation mapping of the human ovary and fallopian tube using optical coherence tomography. J. Biophotonics 16 , e202300002 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Aerts, H. J. et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat. Commun. 5 , 4006 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Pitris, C. et al. High-resolution imaging of gynecologic neoplasms using optical coherence tomography. Obstet. Gynecol. 93 , 135–139 (1999). [ PubMed ] [ Google Scholar ] 29. Ding, B., Jinyuan, T., Tao, K., Ding, Z. & Yang, S. A pilot and ex-vivo study of examination of endometrium tissue by catheter based optical coherence tomography. BMC Med. Imaging 22 , 162 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Law, T. S. M., Wu, F., Xu, H., Wang, C. C. & Li, T. C. Endometrium imaging using real-time rotational optical coherence tomography imaging system. Med. (Baltim.) 98 , e17738 (2019). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Balocco, S., Gatta, C., Pujol, O., Mauri, J. & Radeva, P. SRBF: speckle reducing bilateral filtering. Ultrasound Med. Biol. 36 , 1353–1363 (2010). [ DOI ] [ PubMed ] [ Google Scholar ] 32. Vermeer, K. A., Mo, J., Weda, J. J. A., Lemij, H. G. & de Boer, J. F. Depth-resolved model-based reconstruction of attenuation coefficients in optical coherence tomography. Biomed. Opt. Express 5 , 322–337 (2013). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Otsu, N. A threshold selection method from gray-level histograms. IEEE Trans. Syst., Man, Cybern. 9 , 62–66 (1979). [ Google Scholar ] 34. Maher, J. R. et al. In vivo analysis of burns in a mouse model using spectroscopic optical coherence tomography. Opt. Lett. 39 , 5594–5597 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 35. Haralick, R. M., Shanmugam, K. & Dinstein, I. Textural Features for Image Classification. IEEE Trans. Syst. Man Cybern. SMC-3 , 610–621 (1973). [ Google Scholar ] 36. Yi, J. et al. Spatially resolved optical and ultrastructural properties of colorectal and pancreatic field carcinogenesis observed by inverse spectroscopic optical coherence tomography. J. Biomed. Opt. 19 , 36013 (2014). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Zahed, M. & Skafyan, M. Silhouette-based evaluation of PCA, Isomap, and t-SNE on linear and nonlinear data structures. Stats 8 , 105 (2025). [ Google Scholar ] 38. Nadjafikhah, M. & Nasiri, M. A comparative study of manifold learning methods for scRNA-seq with a trajectory-aware metric. Sci. Rep. 15 , 28923 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Zhao, K. et al. Alzheimer’s disease neuroimaging initiative. Regional radiomics similarity networks reveal distinct subtypes and abnormality patterns in mild cognitive impairment. Adv. Sci. (Weinh.) 9 , e2104538 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Fruchterman, T. M. J. & Reingold, E. M. Graph drawing by force-directed placement. Softw.: Pract. Experience 21 , 1129–1164 (1991). [ Google Scholar ] 41. Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Mathemat. 53–65 (1987). 42. Kaufman, L. & Rousseeuw, P. J. Finding Groups in Data: An Introduction to Cluster Analysis (Wiley, 1990). 43. Board, W. C. “The Female Genital Tumours.” International Agency for Research on Cancer (2020). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary information (8.9MB, docx) Data Availability Statement The patient datasets collected and analyzed for the reported study are not publicly available due to patient privacy concerns; however, the data can be made available upon request with the IRB’s permission. The underlying code to extract functional, structure and radiomic features from OCT data, and code of generating cosine similarity-based network graph are available in GitHub and can be accessed via this link https://github.com/OpticalUltrasoundImaging/OCTEndometrialCancer.git . Articles from npj Imaging are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (10.1 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Assessing Foundation Models for Computational Pathology in Endometrial Cancer#303469
• Exploratory Study of the Interest of MRI Susceptibility Weighted Imaging for the Pre-operative Assessment of Pelvic Endometriosis Extent#1161578
• Prognostic Biomarkers in Patients With Endometrial Cancer#866766
• In vivo verification of treatment source dwell times in brachytherapy of postoperative endometrial carcinoma: a feasibility study#353531
• Screening of Biomarkers on Endometrial Cancers#234439