Evaluating Regions of Clinical Interest Generated by Deep Recurrent Attention Models in Histological Images
Julián Orlando Rodriguez Villamizar, Daniel Felipe Rueda Mariño, David Romo‐Bucheli · 2024
In histological image analysis, an expert patholo-gist might spend a considerable amount of time assessing the anatomical features of the tissues that contain cancerous cells to determine their malignancy. It is common practice to first identify clinically relevant regions at a lower magnification and subsequently analyze the tissue at higher magnification levels. Machine learning techniques offer the potential to streamline these tasks by directly classifying the tissue. However, the interpretation of the results of these models is still an open problem. Our study aims to identify if attended regions in recurrent visual models corresponds to clinical regions of interest defined by experts. Similarly, it is necessary to verify that the model is capable of adequately solving the classification problem to correctly identify images with healthy or cancerous tissues. In our work, we used a recurrent neural network-based image classifier capable of sequentially focus its attention on the most representative areas of histological images to classify histological slides. Even when dealing with a large volume of high-magnification image data, the model can identify relevant regions within the image. We hypothesize that the attended locations should be highly associated with clinically relevant areas in the histological image. We evaluated this hypothesis by measuring the spatial overlapping of the attended locations via probability density divergence metrics, such as Kullback Leibler Divergence, Jensen Shannon Divergence, and Mutual information. The obtained results do not allow us to assert that in general the model's attended regions correspond to clinical regions. However, it was found that for those cases in which the model was able to perform better, in terms of the mentioned metrics, the annotations were more precise and included a more defined area (covering on average 10% of the tissue). Meanwhile, those cases in which a uniform sampling strategy perform better, the annotations covered approximately 25% of the available tissue.