A Fully Automatic Multi-Vendor AI-System To Segment And Predict Resistance To Treatment Of Rectal Cancer On MRI

In this study, we developed and validated a fully automatic system based on pretreatment MRI to predict resistance to therapy in rectal cancer patients using a multi-center and multi-vendor database. Tumors were automatically segmented using in-house automatic U-Net segmentations and subsequently classified as responder and non-responder through a Random Forest algorithm that was fed with a subset of features selected by a customized features selection approach. Despite the strong imbalance between the two classes, the performances yielded are promising, with an area under the curve of 0.72 and a balanced accuracy of 66% on the external validation set. Even if further analyses are still required to improve the performance, our results represent a further step towards a more personalized medicine for patients with rectal cancer.