Advancing Multi-organ and Pan-Cancer Segmentation in Abdominal CT Scans Through Scale-Aware and Self-attentive Modulation

Accurately segmenting abdominal organs and tumors within computed tomography (CT) scans holds paramount significance for facilitating computer-aided diagnosis and devising treatment plans. However, inherent challenges such as lesion heterogeneity and the scarcity of adequately annotated data hamper model development. In this study, we present a two-phase cascaded framework to address the complexities of multi-organ and pan-cancer segmentation. A lightweight CNN first generates candidate regions of interest (ROIs) followed by a hybrid CNN-Transformer model culminating in refined segmentation by synergizing scale-aware modulation for local features and self-attention for global context. Our proposed method secured the 5th position in the MICCAI FLARE23 final test set, showcasing its competitive edge in achieving precise target segmentation with mean Dice Similarity Coefficients of 90.69% for multi-organ and 53.16% for pan-cancer respectively. Additionally, efficient inference is exhibited with an average runtime of 18 s per 512 \(\times \) 512 \(\times \) 215 3D volume with less than 2G GPU memory consumption. Our code is available at: https://github.com/lyupengju/Flare23 .