Medical Transformers for Boosting Automatic Grading of Colon Carcinoma in Histological Images

Developing computer-aided approaches for cancer diagnosis and grading is receiving an uprising demand since this could take over intra- and inter-observer inconsistency, speed up the screening process, allow early diagnosis, and improve the accuracy and consistency of the treatment planning processes. The third most common cancer worldwide and the second most common in women is ColoRectal Cancer (CRC). Grading CRC is a key task in planning appropriate treatments and estimating the response to them. Automatic systems have the potential to speed up and make it more robust but, unfortunately, the most recent and promising machine learning techniques have not been applied for automatic CRC grading so far. For example, there is no work exploiting transformer networks, which outperform convolutional neural networks (CNN) and are replacing them in many applications, for CRC detection and grading at a large scale. To fill this gap, in this work, a transformer-based network endowed with an additional control mechanism in the self-attention module is exploited to understand discriminative regions in large histological images. These relevant regions have been used to train the most suited Convolutional Neural Network (as emerged from recent research findings) for the automatic grading of CRC. The experimental proofs on the largest publicly available CRC dataset demonstrated marked improvement with respect to the leading state-of-the-art approaches relying on CNN