A high-precision fully automated deep learning framework for effective renal segmental artery segmentation in CTA images

Accurate delineation of the perfusion territories of the five renal segmental arteries is essential for renal partial nephrectomy, as it enables reliable identification of distinct renal segments and aids selective clamping and planning. This delineation relies on correct segmentation of the segmental arteries and precision mapping of their respective perfusion territories. For computed tomography angiography (CTA) images, we propose FASM-RSA, a novel fully automated segmentation method for renal segmental arteries. This method divides the renal artery into five segmental arteries: the apical, upper, middle, lower, and posterior segmental arteries. Specifically, the model first replaces consecutive convolutional operations in both the encoder and decoder of the U-shaped architecture with MultiRes modules. This design improves multi-resolution feature extraction and thus better captures renal artery branches across a range of sizes. Second, we design a multi-depth feature aggregation (MDFA) module and a feature selection (FS) module to replace traditional skip connections, which strengthens the model’s ability to extract salient contextual features of the renal arteries. The loss function integrates Dice loss and cross-entropy loss, thereby promoting both overlap-based agreement and pixel-wise prediction accuracy between ground truth and segmentation outputs. Finally, on the CTA184 and CTA405 datasets, the proposed method achieves mean Dice (mDice) scores of 0.782, 0.801, 0.781, 0.876, 0.931 and 0.882, 0.843, 0.836, 0.894, 0.946 for the segmentation of the five renal segmental arteries, respectively. Experimental results demonstrate that our method outperforms state-of-the-art approaches for renal segmental artery segmentation.