A Two-Step Deep Learning Approach for Abdominal Organ Segmentation

Accurate delineation and analysis of anatomical structures within medical images are essential in various clinical applications, with medical image segmentation playing a key role. In the context of abdominal imaging, the precise segmentation of organs like the liver, spleen, and kidneys holds significant importance for tasks such as diagnosis, treatment planning, and surgical interventions. However, achieving precise and efficient segmentation of abdominal organs poses significant challenges due to the variability in organ shape, size, and appearance across different patients and imaging modalities. The MICCAI FLARE23 segmentation paper presents a solution to the challenging problem of segmenting 13 organs and tumor from CT scans, provided 2200 CT scans with partial labels and 1800 CT scans without labels, while balancing model performance and resource consumption. To address these challenges, the paper proposes a two-step segmentation approach that combines organ segmentation and tumor segmentation, which are both accomplished with nnU-Net model. We also crop some top and bottom slices for faster process. Experimental results on the FLARE 2023 test dataset achieved a mean Dice Similarity Coefficient of 0.0361, Normalized Sum of Differences of 0.0331 for organ, a mean Dice Similarity Coefficient of 0.005, Normalized Sum of Differences of 0 for lesion. Besides, our method cost 80.28 s and 158993 MB GPU.