Official Review · ICLR 2024

This work presents a contrastive learning based unsupervised anomaly detection approach. The authors first argue that diffusion models erase fine-grained discriminative features. Based on this assumption the authors add “dissolved samples” (samples with further reverse diffusions) as negative samples to the contrastive framework. These dissolved samples enforce the contrastive learner to differentiate subtleness in the input images. The proposed approach is tested on public datasets, and it demonstrates overall improved performances compared with existing approaches, The problem setting of learning fine-grained subtleness for anomaly detection is of significant clinical potential for medical imaging applications. This paper is well-written with overall sufficient clarity. The insight that diffusion models tend to remove fine-grained discriminative features is interesting. The idea of taking the input images for further reverse diffusions is a bit surpring at first sight but it is reasonable if we assume diffusion models pull reversed samples towards the distributions of input data. A very critical sanity check is needed to validate the dissolving transforms: Would simple gaussian smoothings and/or additive noises on the image space or shallow-layer feature space lead to similar “dissolving” effect as with diffusion models? The general idea is still to remove fine-grained subtleness by smoothing or altering the content. But if simple gaussian smoothing/additive noises can reach similar effects, there would be no need to bother training a diffusion model. This is the major reason I remain slightly negative at this stage. Some terms are not self-explanative: E.g., what does diffusion models to be “feature-aware” mean in the abstract? The term *positive synthesis* and *negative synthesis* may not be sufficiently representative: the authors may consider using the terms *reconstructive approaches* or *generative approaches* to summarize DAE-/AnnoGAN-/Diffusion-based models, etc. and using *discriminative approaches* to summarize synthetic-anomoly-based approaches. Also, the authors may want to discuss approaches that synthesize anomalies based on handcrafted assumptions on anomaly distributions: [1,2]. Performance comparisons with those methods are also welcomed, if the authors find them necessary. The dissolving operation serves as the core of the approach while intuitively it subjects to the selection and training details of the base diffusion models. Therefore, more details of the diffusion model (e.g., subtypes of diffusion models, number of steps, computational overhead, etc.) and how they dictate the final anomaly detection results need to be discussed. Eq. 10 lacks clarity: Do we compute the mean scores under all $k$’s or do we take the max? For $S_{con}$, why does the feature norm need to be multiplied? The authors are encouraged to discuss more on the employed diffusion models, in together with the associated computational cos