From Whole-Body to Abdomen: Streamlined Segmentation of Organs and Tumors via Semi-Supervised Learning and Efficient Coarse-to-Fine Inference

Precise and automated segmentation of abdominal organs and tumors is an important research area of medical image analysis. This domain faces three key challenges: the presence of partially labeled training data that can mislead model training, the variable morphologies of tumors complicating the segmentation process, and the computationally demanding nature of inference in whole/half-body CT scans. In our study, we leverage advanced techniques to generate pseudo-labels, thereby adequately addressing the limitations of partially annotated datasets in a semi-supervised manner. Furthermore, we introduce a novel perspective that allows the segmentation of whole/half-body CT scans to be streamlined into focused abdominal segmentation. To achieve this, we re-engineered the nnU-Net V2 inference engine to incorporate a coarse-to-fine strategy, leading to a remarkable 15 \(\times \) speed-up by eliminating extraneous regions. The mean under the GPU memory-time curve is 7918 Mb. Our approach yields a mean Dice Similarity Coefficient (DSC) of 90.75/47.95 and a Normalized Surface Dice (NSD) of 95.54/40.16 for organ and tumor segmentation, respectively, in the FLARE 2023 validation dataset. Importantly, our method accomplishes these results with an average processing time of only 27.47 s per case.