Response Prediction to Neoadjuvant Chemoradiation in Esophageal Cancer Using Artificial Intelligence & Machine Learning

Overview

In esophageal carcinoma, neoadjuvant concurrent chemo-radiotherapy (NA-CCRT) followed by surgery is the current standard of care and ample evidence has accumulated supporting the view that complete pathological response (pCR) is a positive prognostic marker for improved outcomes. Predicting the probability of achieving pCR prior to neoadjuvant treatment could permit modification of treatment protocols for those patients unlikely to achieve pCR. Radiomics is a new entrant in the field of imaging where specific features are derived from the intensity and distribution pattern of pixels based on a region-of-interest (ROI). The features thus extracted can then be used for prediction modelling similar to other -omics datasets. Preliminary investigations examining its utility have been performed and its applications have thus far focused on screening and survival prediction after treatment. Due to the multi-dimensional nature of data extracted using radiomics, Artificial Intelligence (AI) methods are ideally suited for analysing and modelling radiomic features. Machine Learning (ML) and Deep Learning (DL)\[utilising Convolutional Neural Networks (CNN)\] are both part of the AI framework. In contrast to ML, DL is a new entrant and has been utilised by some medical researchers for modelling using prediction-type algorithms. Besides significantly reducing the workflow associated with Radiomics-based research, feature engineering and modelling using DL are immune to the effects of incorrect ROI delineation. However, the main limitation of DL is the 'blackbox' effect, in which the underlying basis of a CNN is not known. This has been mitigated in part by the visualisation of activation maps directly on the image dataset to prove biological plausibility of predictions. The comparative performance of both types of modelling is also not known. Our objective is to investigate pCR probability in our study population using radiomics-based ML and AI-based modelling. We will also investigate the comparative performance of both modelling techniques. For DL based prediction modelling, we will attempt to provide biological plausibility on the basis of activation maps.

Conditions