3d U-Net with ROI Segmentation of Kidneys and Masses in CT Scans

This project focuses on automatic kidney, tumor and cyst segmentation to assist doctors in diagnosing kidney cancer. We created a deep learning model using methods to first isolate the region of interest of the kidneys, then to segment the kidney and its masses. We used the TotalSegmentator tool to obtain a rough segmentation of the kidneys, then during pre-processing, expanded this region of interest by 18 pixels. This new region of interest was inputted into a 3d segmentation network trained using the nnU-Net library to fully segment the kidneys and masses within them. The current model achieved an average DICE score on validation data of 0.95 for kidney segmentations, and around a 0.8 for tumour and cyst segmentations. On the KiTS23 testing data, the model achieved a 0.94 DICE for kidney segmentations and a 0.73 DICE for mass segmentations.