Accurate Kidney Tumor Segmentation Using Weakly-Supervised Kidney Volume Segmentation in CT images
Mohammad Hossein Sadeghi, Hoda Mohammad Zadeh, Hamid Behroozi, Ali Royat · 2021
Recent leading approaches to medical image segmentation rely on deep convolutional networks trained with human-annotated, pixel-level segmentation labels. However, these methods require training images with pixel-level annotations, which are expensive and time-consuming to obtain. Weakly-supervised approaches have therefore emerged as a solution to address this issue. In CT scan images, due to the large number of slices, pixel-level labeling is very tedious, so applying weakly supervised techniques becomes more necessary. We propose a weakly supervised semantic segmentation approach based on image-level labels for kidney tumor segmentation. Experiments on KITS2019 dataset illustrate that our approach achieves Promising results (Dice score of 0.823 for kidney segmentation and 0.583 for tumor segmentation) with only image-level annotations compared to its fully supervised counterpart in KITS19 challenge (Dice score of 0.974 for kidney segmentation and 0.851 for tumor segmentation).