The effectiveness of an adaptive method to analyse the transition between tumour and peritumour for answering two clinical questions in cancer imaging

Machine Learning (ML) and quantitative cancer imaging have gained huge popularity for contributing to early diagnosis and timeliness of clinical decisions. To increase information of tumour behaviour and progression, several ML studies analyse also the peritumour, and the common approach relies on extending tumour segmentation of a pre-defined fixed size. We present a novel, adaptive, and automatic method to investigate the Zone of Transition (ZoT) bestriding tumour and peritumour, thought as an annular-like shaped area, whose outer and inner borders are detected by analysing gradient variations along tumour borders. Our method is applied to images showing highly different gradient properties: (1)Computed Tomography series of hepatocellular carcinoma for microvascular invasion prediction, (2)Magnetic Resonance series of locally advanced rectal cancer (LARC) for detecting therapy responding patients. ZoT benefits are compared to the common approach, even extracting ML features from gradient magnitude instead of grey level images. As regards HCC, having circular and regular shape, all ML models show similar performance (informedness=0.69, sensitivity=84%, specificity=85%). As regards LARC, with jagged contours, ZoT leads to the best informedness=0.68 (sensitivity=89%, specificity=79%). The marked advantage of our method is detecting the peritumour adaptively, even when not visually noticeable, thus being applicable in different tumours and imaging modalities.