RenalSegNet: automated segmentation of renal tumor, veins, and arteries in contrast-enhanced CT scans

Abstract

Renal carcinoma is a frequently seen cancer globally, with laparoscopic partial nephrectomy (LPN) being the primary form of treatment. Accurately identifying renal structures such as kidneys, tumors, veins, and arteries on CT scans is crucial for optimal surgical preparation and treatment. However, the automatic segmentation of these structures remains challenging due to the kidney's complex anatomy and the variability of imaging data. This study presents RenalSegNet, a novel deep-learning framework for automatically segmenting renal structure in contrast-enhanced CT images. RenalSegNet has an innovative encoder-decoder architecture, including the FlexEncoder Block for efficient multivariate feature extraction and the MedSegPath mechanism for advanced feature distribution and fusion. Evaluated on the KiPA dataset, RenalSegNet achieved remarkable performance, with an average dice score of 86.25%, IOU of 76.75%, Recall of 86.69%, Precision of 86.48%, HD of 15.78 mm, and AVD of 0.79 mm. Ablation studies confirm the critical roles of the MedSegPath and MedFuse components in achieving these results. RenalSegNet's robust performance highlights its potential for clinical applications and offers significant advances in renal cancer treatment by contributing to accurate preoperative planning and postoperative evaluation. Future improvements to model accuracy and applicability will involve integrating advanced techniques, such as unsupervised transformer-based approaches.