Deep Convolutional Neural Networks in Medical Imaging

Dr. Baker’s article about Deep Convolutional Neural Networks (DCNNs) being unable to replicate the configural perception process utilized by humans made me think about what  DCNNs can do and what more they could someday be able to do. Potential uses of Deep Learning Models within neuroscience and biology are frequently related to disease pathology. Utilizing DCNNs has clear implications for recognizing biomarkers of various diseases. Matsuzaka and Iyoda discuss the many implications of convolutional neural networks within diagnostic medicine and pathology. They can analyze a wide range of medical images like x-rays, MRI scans, CT scans, and ultrasound images. They can be more accurate and more efficient than people if the models are trained correctly. Training these algorithms is however a current limitation. It is difficult to compile large and diverse sets of data to train DCNNs on, especially for rare pathology and disease. Dr. Baker and his colleague discussed this problem in their paper and argue that visual models like DCNNs need to be trained on a wide range of object tasks, not just simple category recognition to make them more robust tools. Could training DCNNs to utilize configural perception similarly to human perception help make for more accurate and efficient models when it comes to identifying pathology?

Dr. Baker’s research found that DCNNs are not as accurate as humans when it comes to the tasks that they tested for. However, research that Matsuzaka and Iyoda highlighted a finding of DCNNs performing better than radiologists when it comes to detecting breast cancer in mammograms, leading to fewer false negatives and false positives. If DCNNs were trained to use more configural perception, could they become worse at tasks like recognizing breast cancer in scans? Or would making them use more configural perception allow them to be better utilized across a wider variety of pathologies and could it help with the issue of small sets of data to train models on for less common pathologies?

            Other than the performance aspect of DCNNs in medical image analysis, practicality also has to be considered. There would be a significant price associated with using DCNNs, as data needs to be collected, models need to be trained, and technicians need to be hired to utilize it properly. However it could also lead to less work on the part of people, which would lead to a reduction in the amount of people needed. If they are able to prove more reliable and accurate at analyzing images, there would be less need to extra imaging which would reduce costs.

            Another large concern addressed by Matsuzaka and Iyoda is that of the “black box effect.” Results from DCNNs when it comes to recognition are generally not interpretable leads to the methodology of how the model got to its result not understandable and this lack of transparency leads to mistrust in the models. Research like Dr. Bakers that seeks to elucidate the mechanisms behind DCNNs can help to reduce this effect. By further studying DCNNs and comparing it to human perception, DCNNs usage for medical imaging can be optimized.   

- Sai Kanuru

Baker, N. E., & Elder, J. T. (2022). Deep learning models fail to capture the configural nature of human shape perception. IScience, 25(9), 104913–104913. https://doi.org/10.1016/j.isci.2022.104913

Yasunari Matsuzaka, & Masayuki Iyoda. (2026). Applications, image analysis, and interpretation of computer vision in medical imaging. Frontiers in Radiology, 5, 1733003–1733003. https://doi.org/10.3389/fradi.2025.1733003