Transfer Learning for KiTS21 Challenge

Transfer learning has witnessed a recent surge of interest after proving successful in multiple applications. However, it highly relies on the quantity of annotated data. Constrained by the labor cost and expertise, it is hard to annotate sufficient organs and tumors at the voxel level for medical image segmentation. Consequently, most bench-mark datasets were collected for the segmentation of only one type of organ and/or tumor, and all task-irrelevant organs and tumors were annotated as the background. We aim to make use of these partially but plentifully labeled datasets to boost the segmentation performance of the annotation-limited KiTS21 segmentation task. To this end, we first construct a general medical image segmentation model that learns to segment these partially labeled organs or tumors. Then we transfer its pretrained weights to a specific downstream task, i.e., KiTS21. The primary experiments demonstrate the effectiveness of the proposed transfer learning strategy. Our method achieves 0.890 Dice score, 0.805 SurfaceDice, and 0.822 Tumor Dice in the KiTS21 challenge.

Automated summary

Transfer learning plays a crucial role in medical image segmentation and can improve segmentation performance. By utilizing partially labeled datasets, we can effectively enhance the annotation-limited KiTS21 segmentation task. Experimental results demonstrate the significant success of this method.