Enhanced Segmentation in Abdominal CT Images: Leveraging Hybrid CNN-Transformer Architectures and Compound Loss Function

Accurate segmentation of abdominal organs in CT scans is essential for medical diagnosis and treatment. This paper addresses limitations in current methods by proposing an enhanced HiFormer model for improved segmentation accuracy. We introduce a novel hybrid architecture that combines the strengths of convolutional neural networks (CNNs) and transformers. This model incorporates Cross-covariance image Transformer blocks within the encoder, allowing for efficient spatial information processing. Additionally, a compound DiceTopK loss function optimizes training for better handling organ size variations. This approach effectively addresses the challenges of organ size variability and robustness, surpassing baseline models. Evaluations on the Synapse multi-organ dataset demonstrate significant improvements, achieving a Dice score of 81.15. The proposed method holds promise for enhancing the clinical applications of medical image analysis.