NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1852
Title:Transfusion: Understanding Transfer Learning for Medical Imaging

Reviewer 1

The authors investigate the current transfer learning scheme for deep learning applications to medical imaging. They thoroughly assess and compare the performance of standard architectures that originally designed for the natural image classification tasks with their in-house-developed lightweight and simple models on medical imaging tasks. In this concern, the study demonstrates that latter models can perform comparably with computationally expensive state-of-the-art models. The second finding of the study is that transfer learning does not have a significant benefit for performance. The authors validate the claim by comparing the latent representations of the networks learned with the pretrained weights and training from scratch, and by measuring representational similarity with canonical correlation analysis (CCA). However, in Figure 2, the authors mention that networks from random initialization are more similar representationally to each other. The phenomenon is true for standard pretrained models and smaller versions (CBR-Small and CBR-Tiny) of the proposed models. In contrast, we notice little inconsistencies/mixed performances for two other proposed models (CBR-LargeT and CBR-LargeW). It would be nice if the authors could analyze the situation in the manuscript. The paper illustrates the key properties of transfer learning for medical imaging, analyzing how transfer learning and pretrained weight initialization affect the features and representations learned by the network. The study evidently demonstrates why complex models and pretrained weights are not essential for medical imaging. Also, the work presents feature-independent benefits of transfer learning for better weight scaling and convergence speedups. The motivation of the work is clearly mentioned and explained, and the organization of the study is well-structured. The findings of the study will influence the researchers to develop lightweight deep learning models for the medical imaging domain in upcoming days. However, I am concerned about the originality of the work and significance of the study, considering a high bar to NeurIPS. First of all, the researchers already showed that a simpler network could preserve comparable performance with state-of-the-art networks, and one such study is tuberculosis diagnosis with a CNN [1]. I think this is an incremental contribution to existing work with detailed experiments. Secondly, the study is limited to classification tasks; however, it would be nice if the authors could provide insightful analysis for other medical tasks, such as segmentation, localization, quantification, and so on. Overall, I appreciate the authors’ efforts for their work. One of the main strengths of the study is that with the detailed findings of the paper, the researchers may focus on designing lightweight models for deep learning to medical imaging tasks. Since heavier models significantly impede mobile and on-device applications, the medical domain demands a simpler model in reality. With the limited, but for an overall important contribution, I think the paper marginally above the acceptance threshold. Hence, I vote for accepting the paper. [1] F. Pasa, V. Golkov, F. Pfeiffer, D. Cremers & D. Pfeiffer, ‘Efficient Deep Network Architectures for Fast Chest X-Ray Tuberculosis Screening and Visualization’. Scientific reports, 9(1):6268, 2019. Response to rebuttal: The authors have addressed some of our concerns in their feedback. They have tried to support their hypothesis whether the deep networks and the transfer learning are useful or not in the medical domain through this study. In fact, these are the central questions for many other domains including healthcare, and I have mentioned one of the studies, which shows that simpler network could preserve comparable performance with state-of-the-art networks [1]. However, the authors have not dug the effect of transfer learning explicitly for diverse medical tasks and with the small datasets, while only several hundred to thousand samples are available in many medical tasks. The study is somewhat limited to classification tasks only; however, it would be nice if the authors could provide insightful analysis for other medical tasks, such as segmentation, localization, quantification, and so on. Overall, I still feel that this is an incremental contribution to existing work with more fine-grained experiments based on the carefully chosen dataset. I think this is one of the main drawbacks of the study, which does not make sure about the claim posed in the paper. With such critical limitation, the study could fall short for NeurIPS; however, like other reviewers, I believe that the study is important and should be accessible by the greater community including NeurIPS. Considering all these aspects, I would like to stick to my previous score of 6 but tend to vote for accepting the paper.

Reviewer 2

Overall, I appreciated this paper very much. It addresses the problem of the use and misuse of transfer learning for medical applications and is therefore of high practical value. It is very clearly written and should be accessible to most attendants of NeurIPS. Some minor comments: - Page 2, line 38: microaneurysms appear as small red dots (not yellow). The yellow lesions (which may or may not be roundish) correspond to exudates. - Page 5, line 162: the authors certainly mean Figure 3 (even though Figure 8 can be seen as a confirmation). - Figures 2 and 3: What I found surprising is that the average CCA similarity across layers of the fully trained models is lower than the CCA similarity on the first layer before training, even for random initialization. Probably I missed something, but maybe the authors would like to comment on this, because others might find this intriguing as well. - Page 7, lines 197-199: I do not fully understand the second statement. I do not see from these two graphs that all larger systematically move less than smaller models. For instance, the different CBR variants show nearly the same value on the right plot for all three scenarios, and on the left plot, I also do not see a clear picture. ---- Update: I have read the answers by the authors and found this convincing. The paper is going to make a fine contribution to NeurIPS 2019.

Reviewer 3

I found this paper fascinating to read. I think this paper is quite strong, so my comments will be brief, aside from one recommendation (see Improvements) section that could potentially strengthen this work. Summary: This paper investigates the use of models pretrained on Imagenet that are used for transfer learning on medical applications. This is a ubiquitous practice, and this paper contains many (sometimes counter-intuitive) insights for the field of deep learning on medical images. Primarily, this work suggests that pretraining may be of little to no value for a sufficiently large (see improvements section for more) medical imaging dataset. Even more surprisingly, they find that relatively small models trained solely on the medical data are as good or better than large models like Resnet50 that are pretrained on Imagenet! I find this to be very surprising and exciting. The rest of the paper is devoted to some very creative explorations to explain this key finding. I think this paper should be read by everyone doing medical imagining as it contains numerous pieces of insight and should make us all reconsider what best practices ought to be. Originality Very original and creative paper. They underlying hypothesis (does transfer learning help in medical imaging) Clarity An extremely well written and lucid examination of transfer learning in medical imaging. The presentation flow very smoothly from section to section and is easy to follow. Significance Very high significance not only to medical imaging, but also to computer vision as a whole. Response to rebuttal: I have read the authors responses and found them suitable, and my score of 8 remains