Expansion Rate Parametrization and K-Fold Based Inference with U-Net Neural Networks for Multiclass Medical Image Segmentation

Roman Statkevych, Yuri G. Gordienko, Sergii G. Stirenko

Abstract: In this paper, authors seek a way to improve the performance of state-of-art U-Net models in the context of multiclass segmentation of Human Gastro-Intestinal (GI) Magnetic Resonance Imaging (MRI) by introducing the expansion rate parameter R, which regulates the sizes of layers and allows an increase in the depth of the network while still meeting the model’s size limitations and giving better control over model’s size. For the inference, a method has been used based on k-fold cross-validation which combines an output of each of the k-models. By combining changes in expansion rate, model depth, and using a proposed inference method, it was possible to boost performance by 6% with larger models and reach comparable results with 2x smaller models.