LOSS FUNCTION ANALYSIS FOR CLASS-IMBALANCED MULTI-ORGAN SEGMENTATION OF THE GASTROINTESTINAL TRACT IN MRI

Abstract

MRI guided radiotherapy for the abdominal cancers should must be marked on every scan slice to stomach, small bowel and large bowel so the radiation can avoid healthy tissues. It is often marked by hand, and different experts often outline the same organ differently. While the Deep learning can perform this task automatically, but the data makes it hard to accurate marking. Such as UW-Madison gastrointestinal (GI) tract dataset almost contains 57% no organ and remaining covers only a portion of image when organ appears that leaves the classes heavily imbalanced. The training loss is the main mechanism that drives a network to attend to such rare foreground, yet it is usually chosen by convention rather than by evidence. We compare five losses under identical conditions on a fixed 2.5D network that pairs a SegFormer MiT-B2 encoder with a U-Net decoder: Dice, soft binary cross-entropy (SoftBCE), their combination, Tversky, and a Focal-Dice combination. Training and evaluation use a patient-grouped split and per-image-averaged Dice, intersection over union (IoU), sensitivity, specificity, and precision. All five reach comparable overall Dice within 0.007 (0.9006 to 0.9072), so overall accuracy is largely insensitive to the loss here. The error profile differs sharply, however: Tversky gives the highest sensitivity (0.9465) at the lowest precision (0.9091), SoftBCE the highest precision (0.9363) at the lowest sensitivity (0.9255), and Focal-Dice the best balanced Dice (0.9072). The small bowel stays hardest under every loss. The loss should therefore be chosen for the clinically preferred balance between missing tissue and over-contouring, not for overall accuracy.