A Hounsfiled Unit Grayscale Adaptive Segmentation Network for Kidney Tumor Surgery
Xupeng Kou, Hongcheng Xue, Yakun Yang, Longhe Wang, Li Lin · 2023
Computed tomography (CT)-based structural analysis of the kidney plays a crucial role in kidney tumor surgery. However, existing medical image segmentation algorithms still require improvements due to challenges associated with grayscale conversion and the influence of multi-organ tissues. To address these issues, we propose a multi-label grayscale adaptive segmentation network. Our method incorporates three key innovations: (1) We propose an algorithm to fit original CT data, dynamically adjusting window width via gradient coefficients. This enhances focus on detailed grayscale ranges. (2) We design an feature reorganization module to amalgamate features from diverse sensory fields. This facilitates addressing morphological changes without significantly increasing model parameters. (3) Our intermediate supervision approach enables gradient computation for direct feature extraction from various scales. It supervises model parameter training based on multi-level scaling labels, effectively minimizing computational errors. Experimental results showcase an average Dice similarity coefficient (DSC) surpassing 83.99% in Kipa2022. These outcomes validate the rapid and effective performance of the proposed model, underscoring its clinical significance in kidney tumor treatment.