Deep learning models built from PSMA PET of the primary tumor can predict synchronous and metachronous prostate cancer metastases.

Deep learning models built from PSMA PET of the primary tumor can predict synchronous and metachronous prostate cancer metastases - 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 PLoS One . 2026 Jun 5;21(6):e0349825. doi: 10.1371/journal.pone.0349825 Deep learning models built from PSMA PET of the primary tumor can predict synchronous and metachronous prostate cancer metastases Jesus E Juarez Casillas 1 , Maryam Nezafat 2 , Cecil M Benitez 1 , Kamil Rzechowski 2 , Nathanael Kane 1, 3 , Karl Sjöstrand 2 , Aseem Anand 2 , Ida Sonni 3, 4 , Gholam R Berenji 3 , Sai Duriseti 1, 3 , Matthew B Rettig 5, 6, 7 , Nicholas G Nickols 1, 3, 7, * Editor: Alessandra Castelluccia 8 Author information Article notes Copyright and License information 1 Department of Radiation Oncology, David Geffen School of Medicine at UCLA, Los Angeles, California, United States of America 2 Exini Diagnostics (subsidiary of Lantheus Holdings, Inc), Bedford, Massachusetts, United States of America 3 Department of Radiation Oncology, VA Greater Los Angeles Healthcare System, Los Angeles, CA, United States of America 4 Department of Radiology, David Geffen School of Medicine at UCLA, Los Angeles, California, United States of America 5 Department of Medicine, VA Greater Los Angeles Healthcare System, Los Angeles, California, United States of America 6 Department of Medicine, David Geffen School of Medicine at UCLA, Los Angeles, California, United States of America 7 Department of Urology, David Geffen School of Medicine at UCLA, Los Angeles, California, United States of America 8 Universita Cattolica del Sacro Cuore, ITALY ✉ * E-mail: nicholas.nickols@va.gov , nnickols@mednet.ucla.edu Competing Interests: Dr. Nickols reported receiving grants from Janssen, Lantheus, and Bayer and receiving personal fees from PrimeFour outside the submitted work. Dr. Rettig reported receiving speaking honoraria or consulting fees from Ambrx, Amgen, AVEO, Clovis Oncology, Roivant, Iimmune Bio, Johnson & Johnson, and Bayer outside the submitted work and research funding from Merck, Pfizer, Immune Bio, Novartis, and Fibrogen. Dr Anand, Dr Sjöstrand, Dr Nezafat, and Dr Rzechowski are employed by Exini Diagnostics (subsidiary of Lantheus). No other disclosures were reported. Roles Jesus E Juarez Casillas : Data curation, Formal analysis, Investigation, Methodology, Writing – original draft Maryam Nezafat : Formal analysis, Methodology, Software, Writing – review & editing Cecil M Benitez : Investigation, Writing – review & editing Kamil Rzechowski : Formal analysis, Software, Writing – original draft Nathanael Kane : Data curation, Writing – review & editing Karl Sjöstrand : Investigation, Methodology, Resources, Software, Visualization, Writing – original draft, Writing – review & editing Aseem Anand : Methodology, Resources, Software, Visualization, Writing – review & editing Ida Sonni : Visualization, Writing – review & editing Gholam R Berenji : Visualization, Writing – review & editing Sai Duriseti : Conceptualization, Methodology, Visualization, Writing – review & editing Matthew B Rettig : Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Visualization, Writing – original draft, Writing – review & editing Nicholas G Nickols : Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Visualization, Writing – original draft, Writing – review & editing Alessandra Castelluccia : Editor Received 2025 Aug 29; Accepted 2026 May 5; Collection date 2026. This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. PMC Copyright notice PMCID: PMC13240907  PMID: 42247403 Abstract Objective The objective was to develop prognostic models that included convolutional neural networks (CNN) derived from 18 F-DCFPyL (PSMA) PET imaging of the primary tumor uptake patterns to prognose early metastatic progression after curative intent treatment for localized prostate cancer. Methods Due to the lack of sufficient cases with adequate follow-up and metastatic events to derive this model directly, we derived models that predict the presence of synchronous metastases using only data obtained from the primary tumor. Because early metastatic progression events are consequent to occult metastases present at the time of initial therapy, we hypothesized that a model trained to predict synchronous metastases might also predict metachronous metastatic progression. A convolutional neural network (CNN) model was generated using whole-prostate PSMA PET images and auto-segmented intraprostatic lesions and combined with clinicopathologic data and imaging parameters to develop a multimodal model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) for predicting synchronous metastases and metachronous metastatic progression. Results The multimodal model and CNN model had AUCs of 0.82 (95% CI 0.69–0.92, p < 0.005) and 0.72 (95% CI 0.55–0.84, p 0.0059), respectively, for prediction of synchronous metastases. Shapley additive explanation analysis showed the CNN had the largest contribution to the combined model performance. For metastatic progression, the multimodal model had an AUC of 0.839 (95% CI: 0.6763–1.000, p = 0.0064). Conclusion The multimodal model trained to predict synchronous metastases also predicted metachronous metastatic progression. This supports the potential of artificial intelligence applied to primary tumor PSMA PET images for enhanced prognostication. However, significant limitations include the modest sample size, single center source of data, and potential for overfitting, which may limit generalizability. Therefore, further validation in a larger cohort is indicated. Introduction Prognostic assessments of localized prostate cancer can be derived from conventional clinicopathologic [ 1 , 2 ], transcriptomic [ 3 ], and digital histopathologic data [ 4 ]. In Prostate cancer, AI-based biomarkers and prognostic tools have been developed using data from digitized pathology slides and radiographic images.. Various AI models have been derived using slides from prostate needle biopsies to enhance diagnosis [ 5 ], prognostication [ 4 , 6 , 7 ], and therapeutic selection [ 8 , 9 ]. AI has similarly been applied to diagnostic prostate MRIs to predict recurrence after radical prostatectomy for localized prostate cancer [ 9 – 12 ]. PSMA PET/CT is now routinely used for staging patients with unfavorable intermediate- and high-risk prostate cancer [ 13 ]. AI has been applied to PSMA PET/CT for several purposes including estimation of ISUP grade group [ 14 – 16 ] and to identify [ 17 ], anatomically delineate, and quantify intraprostatic lesions [ 18 , 19 ]. AI tools have also been applied to PSMA PET/CT in the metastatic setting to prognosticate outcome after treatment for metastatic disease [ 20 ]. PSMA PET of the primary tumor may contain additional prognostic information amenable to machine learning to prognosticate outcomes after treatment for localized disease, which is the focus of the present study. Our objective was to develop a model incorporating PSMA PET data from the primary tumor to prognose risk of metastatic progression after curative intent therapy for localized prostate cancer. However, we were limited in outcome data due to the relatively recent availability of PSMA PET imaging at our institution. We hypothesized that a model trained on PSMA PET imaging data within the volume of the prostate to predict synchronous metastatic disease could also identify patients with localized disease on PSMA PET at risk for early metastatic progression. This is hypothesized because most early metastatic progression events after curative intent therapy occur without evidence of local recurrence, suggesting growth of occult metastases present at the time of curative intent therapy are responsible for most of the early metastatic progressions [ 21 ]. We herein report development of a multimodal model derived from a convolutional neural network (CNN) model applied to whole-prostate 18 F-DCFPyL (PSMA) PET images guided by auto-segmented lesions generated by the PSMA PET biomarker platform automated Prostate Cancer Molecular Imaging Standardized Evaluation (aPROMISE) [ 18 , 19 ], combined with clinicopathologic data and measurable imaging parameters that predicts synchronous metastases. We then sought to apply this model to predict metachronous metastases in a cohort of patients with localized prostate cancer after curative intent therapy. Materials and methods PSMA PET scans from 94 treatment naïve patients ( S1 Table ) obtained during initial staging at a single institution with standardized acquisition parameters were used for model development. This retrospective study was approved by the local institutional review board at Veterans Affairs Greater Los Angeles with a waiver of individual informed consent. The imaging showed either unequivocal evidence for metastatic disease (N = 45) or no metastatic disease (N = 49). The non-metastatic cohort subsequently completed curative intent therapy (radical prostatectomy or radiotherapy with androgen deprivation therapy) and had at least 33 months of follow-up without evidence of progression from the date of the scan. aPROMISE is an imaging biomarker platform developed specifically for PSMA PET/CT whole-body quantification and identifies and characterizes findings based on the PROMISE V2 framework [ 22 ]. A core feature of aPROMISE is the segmentation of the body into individual bone structures and other organs, including the prostate. aPROMISE was used to segment the entirety of the prostate and identify intraprostatic lesions while ignoring metastatic lesions. Prostate segmentation on the co-registered CT was used to map the PET image of the prostate. The PET image comprised a three-dimensional cuboid centered on the prostate. The PET image sub-portions containing the entire prostate and corresponding image volumes containing a binary representation of aPROMISE-defined intraprostatic lesions were used as inputs for the CNN ( Fig 1 ). PET imaging artifacts or PET activity outside of the prostate were excluded by aPROMISE. This two 3D channel input (whole-prostate PET image and intraprostatic lesion map) was designed to guide the model’s attention toward areas of the prostate with the cancer while maintaining awareness of the entire prostate volume. Fig 1. Model architecture. Open in a new tab The DCFPyL PET images are first segmented using aPROMISE to delineate the prostate. A fixed-size 3D cuboid centered on the segmented prostate with the corresponding binary representation of the intraprostatic lesions is extracted and used as a two-channel input to a modified 3D SqueezeNetv2 mode (any metastases present were ignored). The output of the CNN -only model is obtained from the final fully connected layer as a single continuous score. For multimodal prediction, this CNN score is combined with clinical features using a Naïve Bayes classifier to generate the final multimodal output score. The multimodal score and CNN score had AUCs of 0.82 (95% CI 0.69-0.92, p < 0.005) and 0.72 (95% CI 0.55-0.84, p 0.0059), respectively, for prediction of synchronous metastases. Several different CNN architectures were considered. Because of the limited number of training samples and the high risk of model overfitting, only those with a low number of parameters were considered, including ResNet18 [ 23 ], SqueezeNetv2, [ 24 ] MobileNet, [ 25 ] and ShuffleNet [ 26 ]. Squeezenetv2 was selected due to its compactness and associated low computational requirements. The architecture was adapted for volumetric image analysis by using 3D convolutions and pooling. All convolutional operations were implemented using 3D kernels of size 1x1x1 and 3x3x3 with padding applied to preserve spatial dimensions. For the CNN input, a cubic region of interest centered on the prostate area was extracted from the PET scan. Each input volume had a fixed size of 14x54x54 corresponding to the prostate area within the pelvic region. Images were normalized on a per-volume basis using min-max normalization to the range of [0,1]. Two input channels were used, the PET scan and corresponding hotspot representation. These channels were concatenated along the input channel dimension. PET and hotspots were concatenated along the channel dimension and fed to the CNN model. The base SqueezeNet architecture is described in [ 21 ]. The original SqueezeNet classifier (last Dropout, Convolution with kernel size 1, ReLU activation, and Average 3D Pooling layers) was replaced by Adaptive Max Pooling 3D (Output Size 1), Dropout (p = 0.8) and a Fully Connected layer mapping 512 features to a single output. The modified 3D SqueezeNet retained a lightweight design with approximately 1 million trainable parameters. Models were trained using the weighted Binary Cross Entropy loss function. The model was further optimized by performing grid-search hyperparameter-tuning. The optimal values for group convolution, learning rate, regularization strength, and set of augmentations were determined based on initial validation. We used a learning rate of 1e-4 with ReduceLROnPlateau scheduling. The neural network was trained for maximum of 300 epochs using early stopping with a patience of 140 epochs and the checkpoint for the best-performing epoch on the evaluation subset was kept for subsequent testing. To avoid overfitting, a randomly selected set of augmentations was applied during training. At each epoch, a random subset of seven out of fifteen augmentations were selected and applied in random order exclusively to the training data. The augmentation pool included: random rotation, random flip in left-right axes, random Gaussian noise, random standard deviation intensity shift, random contrast adjustment, random Gaussian smooth, random Gaussian sharpen, random histogram shift, random Coarse shuffle, random 3D elastic distortion, random affine transformation, random Gibbs noise, random bias field, random K-space Spike Noise, and random Rician Noise. Random rotations were limited to +/-20◦ to preserve the prostate orientation as the model architecture consists of 3x3x3 convolutional kernels, which from design are not rotation invariant. We also developed a combined (multimodal) model that added conventional clinicopathologic data (PSA, pathologic grade group, percent positive cores) and measurable imaging parameters (peak SUV prostate-located value, PRIMARY score, PSMA expression score per PROMISE-V2) [ 27 , 28 ] to the CNN via a Naïve Bayes classifier ( Fig 1 ). A Naïve Bayes classifier was selected over other alternatives for its robustness in small, high-dimensional datasets, and compatibility with CNN-derived and clinical features. The dataset was split into training, validation, and test sets using stratified random sampling in ratios of 0.5, 0.2, and 0.3. In total, 94 patients ( S1 Table ) were included (Training: 47, validation:19 and test: 28). Model performance in predicting synchronous metastases was evaluated using the area under the receiver operating characteristic curve (AUC), with imaging-based ground truth as reference. Model training was repeated across 50 independent runs, and the best-performing model, as determined by AUC on the validation set was selected for final evaluation. The final CNN model and multimodal model were then tested to predict metachronous metastases in a cohort of patients with localized prostate cancer after curative intent therapy ( S2 Table ). Imaging and clinicopathologic data used in this project were retrospectively collected between May 4, 2023 and February 10, 2024. The authors that retrospectively accessed the data had access to information that could identify individual patients during data collection. Results and discussion The CNN and multimodal models achieved AUCs of 0.82 (95% CI 0.69–0.92, p < 0.005) and 0.72 (95% CI 0.55–0.84, p 0.0059), respectively, for synchronous metastases prediction ( Fig 1 ). When considering each input independently, the CNN was the greatest contributing input based on Shapley additive explanation (SHAP) analysis ( Fig 2A ). Image based attention maps were generated using the FoxAI framework that demonstrates the model focuses attention on the regions of tumor within the prostate ( S1 Fig ). The models were applied to a cohort (n = 23) who had localized disease at initial PSMA PET, underwent curative intent therapy (radical prostatectomy or radiotherapy with androgen deprivation therapy), had at least four years follow-up, and had either no evidence of progression up to four years (n = 13), or metastatic progression by PSMA PET within four years (n = 10). Performance of metachronous metastases prediction was evaluated using receiver operating characteristics (ROC) analysis with discrimination quantified by the area under the ROC curve (AUC) and 95% confidence intervals calculated using DeLong’s method. The multimodal model achieved an AUC of 0.839 (95% CI: 0.6763–1.000, p = 0.0064), outperforming the CNN only model AUC of 0.723 (95% CI: 0.497–0.949, p = 0.072), the CAPRA score AUC of 0.742 (95% CI:0.55–0.934, p = 0.051) and UCLA PSMA Risk Score AUC of 0.677 (95% CI:0.457–0.896, p = 0.1538) ( Fig 2B - 2C ). In summary, the CNN only model performed similarly to the CAPRA and UCLA PSMA Risk Score models that were comprised entirely of clinical features. The multimodal model, comprised of both the image-based CNN as well as clinical features, performed modestly better. Kaplan-Meier plots at tertiles and the median demonstrate differences in metastatic progression based on the multimodal score ( Fig 3 ). Fig 2. Performance of the multimodal model. Open in a new tab A. Feature importance derived from Shapley additive explanation (SHAP) analysis showing the impact of contribution in descending order: CNN score, PSA value, pathologic grade group, percent positive cores, PRIMARY score, PSMA expression score, and peak PSMA uptake value. B. Receiver Operating Characteristic (ROC) curve of the multimodal model, with the area under the curve (AUC), C. Comparison of model performance, reported as AUC with 95% confidence intervals and corresponding p-values. Fig 3. Kaplan-Meier survival analysis. Open in a new tab Kaplan-Meier survival curves by multimodal score per tertile ( A ) and by median (B) . Our objective was to develop a prognostic model for early metastatic progression after curative intent therapy derived from PSMA PET data in the prostate. However, we lacked sufficient patients with adequate follow-up and metastatic progression events to train such a model directly. Instead, we developed a model that could predict for synchronous metastases by evaluating the PSMA PET data in the primary tumor. In training the model, non-metastatic cases were purposefully selected to minimize contamination from scans of patients who had early metastatic progression events and the metastatic cases were purposefully selected to minimize risk of false positives. Furthermore, a subset of patients with actual metachronous progression was withheld from model training and reserved for final model evaluation. The multimodal model trained to predict synchronous metastases also predicted metachronous metastatic progression. This is consistent with the notion that occult metastases present at the time of initial curative intent therapy account for early metastatic progressions [ 21 ]. There are several limitations. One limitation to our study is that all scans used to develop the model were from a single center. This may limit generalizability. Moreover, the moderate size in the initial dataset resulted in a limited quantity of training samples, which limits model performance and narrows the options for model architecture. Given the case numbers used for model development, it is possible that reported AUCs are overestimates. Overfitting is another potential limitation. To address this, we employed models with a low number of parameters and implemented a random selection of augmentations for the CNN input. Another limitation is the small cohort used to test the model for metachronous progression which was severely limited by our access to images of patients who underwent PSMA PET with sufficient follow-up post treatment. These limitations render the findings as hypothesis generating and preliminary; the study is best interpreted as proof of concept. Accordingly, further validation in a larger multicenter cohort with additional performance metrics is necessary. A combined model of similar architecture but trained on a larger data set with actual cases of metastatic progression may have greater prognostic value. Supporting information S1 Table. Clinical pathologic features corresponding to the images that were used as inputs to train the model. Four scans did not have intraprostatic lesions identifiable by aPROMISE and were not used in the final CNN model. UIR = unfavorable intermediate risk, HR = high risk. *Diagnosed on metastatic biopsy and had no prostate biopsy. (PDF) pone.0349825.s001.pdf (145.4KB, pdf) S2 Table. Clinical pathologic features of cases used to test the multimodal model prediction of metachronous recurrence. UIR = unfavorable intermediate risk, HR = high risk. (PDF) pone.0349825.s002.pdf (142.8KB, pdf) S1 Fig. Attention maps. A . PET prostate-area cube. B . Intraprostatic lesions identified by aPROMISE. C . Attention maps using FoXAI. (PDF) pone.0349825.s003.pdf (76.1KB, pdf) Data Availability Individually identifiable private information, protected health information (PHI), and other sensitive data will never be shared. Only de-identified data may be shared. Requests may be made to the VA Greater Los Angeles Healthcare System Research Service (11301 Wilshire Blvd, Los Angeles, CA 90073) for researchers who meet the criteria for access to the data through a data sharing agreement. Funding Statement This study was funded through a Cooperative Research and Development Agreement between VA Greater Los Angeles and Lantheus. The funder provided support in the form of salaries for authors MN, KR, KS, AA, and these authors contributed to the development of the code used for the CNN, and multimodal model. The specific roles of these authors are articulated in the ‘author contributions’ section. References 1. Cooperberg MR, Pasta DJ, Elkin EP, Litwin MS, Latini DM, Du Chane J. The University of California, San Francisco cancer of the prostate risk assessment score: a straightforward and reliable preoperative predictor of disease recurrence after radical prostatectomy. J Urol. 2005;173:1938–42. doi: 10.1097/01.ju.0000158155.33890.e7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Xiang M, Ma TM, Savjani R, Pollom EL, Karnes RJ, Grogan T, et al. Performance of a prostate-specific membrane antigen positron emission tomography/computed tomography-derived risk-stratification tool for high-risk and very high-risk prostate cancer. JAMA Netw Open. 2021;4(12):e2138550. doi: 10.1001/jamanetworkopen.2021.38550 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Erho N, Crisan A, Vergara IA, Mitra AP, Ghadessi M, Buerki C, et al. Discovery and validation of a prostate cancer genomic classifier that predicts early metastasis following radical prostatectomy. PLoS One. 2013;8(6):e66855. doi: 10.1371/journal.pone.0066855 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Esteva A, Feng J, van der Wal D, Huang S-C, Simko JP, DeVries S, et al. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. NPJ Digit Med. 2022;5(1):71. doi: 10.1038/s41746-022-00613-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Bulten W, Kartasalo K, Chen P-HC, Ström P, Pinckaers H, Nagpal K, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154–63. doi: 10.1038/s41591-021-01620-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Li W, Li J, Sarma KV, Ho KC, Shen S, Knudsen BS, et al. Path R-CNN for prostate cancer diagnosis and gleason grading of histological images. IEEE Trans Med Imaging. 2019;38(4):945–54. doi: 10.1109/TMI.2018.2875868 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Leo P, Janowczyk A, Elliott R, Janaki N, Bera K, Shiradkar R, et al. Computer extracted gland features from H&E predicts prostate cancer recurrence comparably to a genomic companion diagnostic test: a large multi-site study. NPJ Precis Oncol. 2021;5(1):35. doi: 10.1038/s41698-021-00174-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Spratt DE, Tang S, Sun Y, Huang H-C, Chen E, Mohamad O, et al. Artificial Intelligence Predictive Model for Hormone Therapy Use in Prostate Cancer. NEJM Evid. 2023;2(8):EVIDoa2300023. doi: 10.1056/EVIDoa2300023 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. 2022;19(2):132–46. doi: 10.1038/s41571-021-00560-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Hiremath A, Shiradkar R, Fu P, Mahran A, Rastinehad AR, Tewari A, et al. An integrated nomogram combining deep learning, Prostate Imaging-Reporting and Data System (PI-RADS) scoring, and clinical variables for identification of clinically significant prostate cancer on biparametric MRI: a retrospective multicentre study. Lancet Digit Health. 2021;3(7):e445–54. doi: 10.1016/S2589-7500(21)00082-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Li L, Shiradkar R, Leo P, Algohary A, Fu P, Tirumani SH, et al. A novel imaging based Nomogram for predicting post-surgical biochemical recurrence and adverse pathology of prostate cancer from pre-operative bi-parametric MRI. EBioMedicine. 2021;63:103163. doi: 10.1016/j.ebiom.2020.103163 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Hiremath A, Corredor G, Li L, Leo P, Magi-Galluzzi C, Elliott R, et al. An integrated radiology-pathology machine learning classifier for outcome prediction following radical prostatectomy: preliminary findings. Heliyon. 2024;10(8):e29602. doi: 10.1016/j.heliyon.2024.e29602 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Hofman MS, Lawrentschuk N, Francis RJ, Tang C, Vela I, Thomas P, et al. Prostate-specific membrane antigen PET-CT in patients with high-risk prostate cancer before curative-intent surgery or radiotherapy (proPSMA): a prospective, randomised, multicentre study. Lancet. 2020;395(10231):1208–16. doi: 10.1016/S0140-6736(20)30314-7 [ DOI ] [ PubMed ] [ Google Scholar ] 14. Liu J, Cundy TP, Woon DTS, Desai N, Palaniswami M, Lawrentschuk N. A systematic review on artificial intelligence evaluating PSMA PET scan for intraprostatic cancer. BJU Int. 2024;134(5):714–22. doi: 10.1111/bju.16412 [ DOI ] [ PubMed ] [ Google Scholar ] 15. Yao F, Bian S, Zhu D, Yuan Y, Pan K, Pan Z, et al. Machine learning-based radiomics for multiple primary prostate cancer biological characteristics prediction with 18F-PSMA-1007 PET: comparison among different volume segmentation thresholds. Radiol Med. 2022;127(10):1170–8. doi: 10.1007/s11547-022-01541-1 [ DOI ] [ PubMed ] [ Google Scholar ] 16. Ghezzo S, Mapelli P, Bezzi C, Samanes Gajate AM, Brembilla G, Gotuzzo I, et al. Role of [68Ga]Ga-PSMA-11 PET radiomics to predict post-surgical ISUP grade in primary prostate cancer. Eur J Nucl Med Mol Imaging. 2023;50(8):2548–60. doi: 10.1007/s00259-023-06187-3 [ DOI ] [ PubMed ] [ Google Scholar ] 17. Holzschuh JC, Mix M, Ruf J, Hölscher T, Kotzerke J, Vrachimis A, et al. Deep learning based automated delineation of the intraprostatic gross tumour volume in PSMA-PET for patients with primary prostate cancer. Radiother Oncol. 2023;188:109774. doi: 10.1016/j.radonc.2023.109774 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Nickols N, Anand A, Johnsson K, Brynolfsson J, Borreli P, Parikh N, et al. aPROMISE: a novel automated PROMISE platform to standardize evaluation of tumor burden in 18F-DCFPyL images of veterans with prostate cancer. J Nucl Med. 2022;63(2):233–9. doi: 10.2967/jnumed.120.261863 [ DOI ] [ PubMed ] [ Google Scholar ] 19. Johnsson K, Brynolfsson J, Sahlstedt H, Nickols NG, Rettig M, Probst S, et al. Analytical performance of aPROMISE: automated anatomic contextualization, detection, and quantification of [18F]DCFPyL (PSMA) imaging for standardized reporting. Eur J Nucl Med Mol Imaging. 2022;49(3):1041–51. doi: 10.1007/s00259-021-05497-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Zang S, Meng Q, Li X, Guo T, Zhang L, Zhao Z, et al. Fully automated volumetric assessment of tumor burden using artificial intelligence on 68Ga-PSMA-11 PET predicts survival after 177Lu-PSMA therapy in metastatic Castration-resistant prostate cancer. Eur J Nucl Med Mol Imaging. 2026;53(3):1913–26. doi: 10.1007/s00259-025-07628-x [ DOI ] [ PubMed ] [ Google Scholar ] 21. Ma TM, Chu F-I, Sandler H, Feng FY, Efstathiou JA, Jones CU, et al. Local failure events in prostate cancer treated with radiotherapy: a pooled analysis of 18 randomized trials from the meta-analysis of randomized trials in cancer of the prostate consortium (LEVIATHAN). Eur Urol. 2022;82(5):487–98. doi: 10.1016/j.eururo.2022.07.011 [ DOI ] [ PubMed ] [ Google Scholar ] 22. Seifert R, Emmett L, Rowe SP, Herrmann K, Hadaschik B, Calais J, et al. Second version of the prostate cancer molecular imaging standardized evaluation framework including response evaluation for clinical trials (PROMISE V2 ). Eur Urol. 2023;83: 405–12. doi: 10.1016/j.eururo.2023.02.002 [ DOI ] [ PubMed ] [ Google Scholar ] 23. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: 2016 IEEE Conference on computer vision and pattern recognition (CVPR), 2016. 770–8. doi: 10.1109/cvpr.2016.90 [ DOI ] [ Google Scholar ] 24. Iandola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size. arXiv. 2016. doi: 10.48550/arXiv.1602.07360 [ DOI ] [ Google Scholar ] 25. Howard AG, Zhu M, Chen B, Kalenichenko D, Wang W, Weyand T. MobileNets: efficient convolutional neural networks for mobile vision applications. 2017. doi: 10.48550/arXiv.1704.04861 [ DOI ] [ Google Scholar ] 26. Zhang X, Zhou X, Lin M, Sun J. ShuffleNet: an extremely efficient convolutional neural network for mobile devices. arXiv. 2017. doi: 10.48550/arXiv.1707.01083 [ DOI ] [ Google Scholar ] 27. Emmett L, Papa N, Buteau J, Ho B, Liu V, Roberts M, et al. The PRIMARY score: using intraprostatic 68Ga-PSMA PET/CT patterns to optimize prostate cancer diagnosis. J Nucl Med. 2022;63(11):1644–50. doi: 10.2967/jnumed.121.263448 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Seifert R, Emmett L, Rowe SP, Herrmann K, Hadaschik B, Calais J, et al. Second version of the prostate cancer molecular imaging standardized evaluation framework including response evaluation for clinical trials (PROMISE V2 ). Eur Urol. 2023;83: 405–12. doi: 10.1016/j.eururo.2023.02.002 [ DOI ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials S1 Table. Clinical pathologic features corresponding to the images that were used as inputs to train the model. Four scans did not have intraprostatic lesions identifiable by aPROMISE and were not used in the final CNN model. UIR = unfavorable intermediate risk, HR = high risk. *Diagnosed on metastatic biopsy and had no prostate biopsy. (PDF) pone.0349825.s001.pdf (145.4KB, pdf) S2 Table. Clinical pathologic features of cases used to test the multimodal model prediction of metachronous recurrence. UIR = unfavorable intermediate risk, HR = high risk. (PDF) pone.0349825.s002.pdf (142.8KB, pdf) S1 Fig. Attention maps. A . PET prostate-area cube. B . Intraprostatic lesions identified by aPROMISE. C . Attention maps using FoXAI. (PDF) pone.0349825.s003.pdf (76.1KB, pdf) Data Availability Statement Individually identifiable private information, protected health information (PHI), and other sensitive data will never be shared. Only de-identified data may be shared. Requests may be made to the VA Greater Los Angeles Healthcare System Research Service (11301 Wilshire Blvd, Los Angeles, CA 90073) for researchers who meet the criteria for access to the data through a data sharing agreement. Articles from PLOS One are provided here courtesy of PLOS ACTIONS View on publisher site PDF (617.9 KB) 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

• Deep learning models built from PSMA PET of the primary tumor can predict synchronous and metachronous prostate cancer metastases#265686
• A deep learning based NeuroFusionNet approach for automated brain tumor diagnosis from MRI.#150302
• Deep Learning Applications in Medical Imaging#491026
• Machine and deep learning based on magnetic resonance imaging to segment glioblastoma and predict the spread of recurrence: a multicenter retrospective protocol.#402599
• Performance of MRI-based deep learning models in differentiation of triple negative breast cancer from other breast cancer subtypes: A systematic review and meta-analysis.#663319