For many years, convolutional neural networks dominated the field of computer vision, not least in the medical field, where problems such as image segmentation were addressed by such networks as the U-Net. The arrival of self-attention-based networks to the field of computer vision through ViTs seems to have changed the trend of using standard convolutions. Throughout this work, we apply different architectures such as U-Net, ViTs and ConvMixer, to compare their performance on a medical semantic segmentation problem. All the models have been trained from scratch on the DRIVE dataset and evaluated on their private counterparts to assess which of the models performed better in the segmentation problem. Our major contribution is showing that the best-performing model (ConvMixer) is the one that shares the approach from the ViT (processing images as patches) while maintaining the foundational blocks (convolutions) from the U-Net. This mixture does not only produce better results ($DICE=0.83$) than both ViTs ($0.80$/0.077 for UNETR/SWIN-Unet) and the U-Net ($0.82$) on their own but reduces considerably the number of parameters (2.97M against 104M/27M and 31M, respectively), showing that there is no need to systematically use large models for solving image problems where smaller architectures with the optimal pieces can get better results.

There is an open debate on the role of artificial networks to understand the visual brain. Internal representations of images in artificial networks develop human-like properties. In particular, evaluating distortions using differences between internal features is correlated to human perception of distortion. However, the origins of this correlation are not well understood. Here, we dissect the different factors involved in the emergence of human-like behavior: function, architecture, and environment. To do so, we evaluate the aforementioned human-network correlation at different depths of 46 pre-trained model configurations that include no psycho-visual information. The results show that most of the models correlate better with human opinion than SSIM (a de-facto standard in subjective image quality). Moreover, some models are better than state-of-the-art networks specifically tuned for the application (LPIPS, DISTS). Regarding the function, supervised classification leads to nets that correlate better with humans than the explored models for self- and non-supervised tasks. However, we found that better performance in the task does not imply more human behavior. Regarding the architecture, simpler models correlate better with humans than very deep nets and generally, the highest correlation is not achieved in the last layer. Finally, regarding the environment, training with large natural datasets leads to bigger correlations than training in smaller databases with restricted content, as expected. We also found that the best classification models are not the best for predicting human distances. In the general debate about understanding human vision, our empirical findings imply that explanations have not to be focused on a single abstraction level, but all function, architecture, and environment are relevant.

The experiments allowed by current machine learning models imply a revival of the debate on the causes of specific trends of human visual psychophysics. Machine learning facilitates the exploration of the effect of specific visual goals (such as image segmentation) by different neural architectures in different statistical environments in an unprecedented manner. In this way, (1) the principles behind psychophysical facts such as the non-Euclidean nature of human color discrimination and (2) the emergence of human-like behaviour in artificial systems can be explored under a new light. In this work, we show for the first time that the tolerance or invariance of image segmentation networks for natural images under changes of illuminant in the color space (a sort of insensitivity region around the white) is an ellipsoid oriented similarly to a (human) MacAdam ellipse. This striking similarity between an artificial system and human vision motivates a set of experiments checking the relevance of the statistical environment on the emergence of such insensitivity regions. Results suggest, that in this case, the statistics of the environment may be more relevant than the architecture selected to perform the image segmentation.

One of the key problems in computer vision is adaptation: models are too rigid to follow the variability of the inputs. The canonical computation that explains adaptation in sensory neuroscience is divisive normalization, and it has appealing effects on image manifolds. In this work we show that including divisive normalization in current deep networks makes them more invariant to non-informative changes in the images. In particular, we focus on U-Net architectures for image segmentation. Experiments show that the inclusion of divisive normalization in the U-Net architecture leads to better segmentation results with respect to conventional U-Net. The gain increases steadily when dealing with images acquired in bad weather conditions. In addition to the results on the Cityscapes and Foggy Cityscapes datasets, we explain these advantages through visualization of the responses: the equalization induced by the divisive normalization leads to more invariant features to local changes in contrast and illumination.