Robustness of machine learning techniques for CT26 colon carcinoma classification based on hyperspectral images

Cancer is a significant cause of death, affecting millions worldwide. Urgent attention is required to detect cancer early and monitor its development. As tumors exhibit altered physiology and morphology compared to healthy tissues, they can be non-invasively detected using optical imaging techniques.

In this study, we utilized hyperspectral imaging (HSI) to monitor CT26 colon carcinoma tumors subcutaneously grown on the back of BALB/c mice. We extracted physiological properties such as total hemoglobin and tissue oxygenation using the inverse adding-doubling (IAD) algorithm. Then, we selected the ten most relevant physiological and morphological features describing tumors and utilized 30 advanced machine learning (ML) algorithms to discriminate tumors from healthy tissues. Ultimately, we tested the robustness of feature selection and model prediction by variably smoothing tissue parameter images and systematically omitting training data.

Average AUC saw a 0.1 increase when employing smoothing up to sigma = 1.5 and omitting training data from Days 1-3 and 9 of the experiment. We achieved a maximum AUC of 1 in distinguishing CT26 tumors from healthy tissues. Our findings underscore the significance of combining novel optical imaging techniques with cutting-edge ML approaches for tumor detection.