Hybrid Swin Transformer-CNN segmentation of muscle and adipose tissues from CT images for sarcopenia assessment in patients with borderline resectable pancreatic cancer

Background: This study aimed to develop and validate a hybrid Swin Transformer-CNN model for automated CT-based segmentation of muscle and adipose tissues at the L3 level and to evaluate its ability to stratify survival in patients with borderline resectable pancreatic cancer (BRPC). Methods: This retrospective multicenter study included 110 BRPC patients evaluated between 2015-2023. A total of 330 axial CT slices at L3 were extracted. An initial CNN-RNN pipeline automatically selected L3 slices. A hybrid Swin Transformer-CNN segmentation architecture was trained to delineate skeletal muscle and adipose tissues. Internal validation used 66 slices. Independent expert validation involved two senior radiologists segmenting 111 additional slices to assess inter-observer variability and establish a consensus reference. The model’s performance was compared against this consensus and other published segmentation architectures. Clinical relevance was assessed by calculating skeletal muscle index (SMI) and stratifying overall survival (OS) and progression-free survival (PFS) using established sarcopenia thresholds. Kaplan-Meier estimates and log-rank tests were used for statistical analysis. Results: The CNN-RNN achieved 97.3% accuracy for L3 slice selection. The model outperformed traditional architectures, with a mean Dice score of 0.97 for skeletal muscle (IoU: 0.88). Against expert consensus, it achieved a Dice of 0.94 (±0.016); inter-observer variability was 0.975. Automated SMI identified 10 sarcopenic patients with significantly reduced OS (16 vs. 26 months, p=0.02); PFS differences were not significant (12 vs. 16 months, p=0.14). Conclusion: This model enables accurate, automated CT muscle and adipose segmentation, replicates expert annotations, and supports sarcopenia-based survival stratification in BRPC, aiding personalized oncology decisions.