Artificial Intelligence-based prediction of treatment response to neoadjuvant chemoradiotherapy in colorectal cancer patients using digital pathology.

Colorectal cancer (CRC) is the third most common cancer in humans, and affects mainly people from developed countries. This cancer arises from polyps, which may turn into tumours after growing for 10-20 years. The diagnosis of CRC is performed through examination of a sample obtained from a biopsy. Currently, CRCs are treated through surgical resection, often preceded by neoadjuvant chemoradiotherapy, an intervention aimed at reducing tumour size before the surgical procedure. However, only 30% of patients affected by CRC achieve a complete pathological response to this treatment. In the context of CRC diagnosis, Digital Pathology (DP) is a new and innovative research field that assists the pathologists by enabling the acquisition, visualisation and processing of histological images. Specifically, biopsy specimens are manually processed and prepared for digitisation and visualisation. As the result of this preparation protocol, a high-resolution virtual slide is obtained, known as a Whole Slide Image (WSI). Although the visual examination of a WSI by an expert pathologist is still the standard for tumour diagnosis, this is a challenging and laborious task that is subject to inter- and intra-operator variability. Consequently, several automated algorithms have been proposed for the analysis of WSIs, supported by technological advances in recent years. In this regard, most of these methods focus on the detection and segmentation of the tumour region within a histological image, while the context of treatment response prediction is not so well addressed. Based on these premises, the aim of this study is to provide a Computer Aided Diagnosis (CAD) system for predicting treatment response through the analysis of histological images. Specifically, the treatment response is classified into two different categories: "Resistant" is used for patients who did not achieve a complete pathological response to the chemoradiotherapy, while "Sensitive" is used for patients who had a favorable response. Therefore, this study applies Artificial Intelligence (AI) algorithms to the analysis of WSIs and examines the difficulties and challenges associated with the management of this type of data, using two different approaches. The first strategy involves a Deep Learning (DL) classification network to predict treatment response at the patch level. Each histological image, in this instance, is processed at its highest resolution by dividing it into tiles. Furthermore, certain selection steps are implemented to handle and reduce the amount of data and to adapt it to the DL model. The network outputs are then aggregated to yield the final prediction of the treatment response at the WSI level that characterises each subject. In the second approach, histological images are processed at a lower resolution level than the previous one. In particular, radiomics features are extracted from this level and used as inputs for Machine Learning (ML) classifiers. Finally, it is examined whether the integration of the probabilities of belonging to each of the two classes for each patch of a subject, obtained from the previous DL model, could be an additional feature. In conclusion, this study introduces a CAD system that utilises Artificial Intelligence to predict treatment responses in CRC. By combining DL and ML approaches, the system analyses analyses histological images at different levels of magnification to suggest which provide the most relevant information about the pathological response to therapy.