Abstract: Using the CNN technique, the TKP model is developed to automatically present the renal artery trees and precisely delineate the perfusion regions of different segmental arteries. The guidance of the TKP model is feasible and effective in nephron-sparing surgery.
Introduction
Nephron-sparing surgery (NSS), or partial nephrectomy, is the clinical gold standard for treating localized kidney tumors. A primary challenge during this procedure is managing blood flow to prevent excessive bleeding. Traditionally, surgeons employ "global ischemia" by clamping the main renal artery, which stops blood flow to the entire kidney. However, this approach risks permanent warm ischemic injury to healthy tissue.
Segmental Renal Artery Clamping (SRAC) offers a more targeted alternative by clamping only the specific branches feeding the tumor. Despite its benefits, SRAC is difficult to implement because renal arterial anatomy is highly complex and varies significantly between patients. Manual planning using CT scans is slow and often fails to accurately map the specific "perfusion territory" of each arterial branch. To address this, researchers developed the Tridimensional Kidney Perfusion (TKP) model, an automated 3D deep learning pipeline designed to create a precise "digital twin" of the patient's renal vasculature.
Methodology: The TKP Pipeline
The TKP model utilizes a multi-stage deep learning approach to automate surgical planning:
1. Kidney and Tumor Segmentation
The first stage involves the precise delineation of the kidney and the tumor using a 3D Fully Convolutional Network integrated with a Pyramid Pooling Module (3D_FCN_PPM). This architecture is specifically designed to capture both local fine-grained details and global anatomical context.
The 3D_FCN_PPM network employs residual blocks to facilitate deep feature learning and a Pyramid Pooling Module to aggregate multi-scale information, ensuring robust segmentation across varying tumor sizes and locations.
2. Artery Segmentation and Perfusion Mapping
Following kidney segmentation, the DPA-DenseBiasNet is used for arterial mapping. This stage leverages "Deep-priori anatomy" (DPA)—an unsupervised autoencoder trained on large unlabeled datasets to learn general kidney structures. These priors guide the supervised segmentation of small, complex arterial branches. Finally, a distance transformation algorithm assigns every kidney voxel to its nearest arterial branch, generating a color-coded perfusion map that predicts exactly which sections of the kidney are supplied by specific arteries.
Validation and Clinical Results
The study utilized several datasets for development and validation, including CT images from 140 patients for the kidney model and 196 patients for the arterial model. Validation was conducted through both porcine and clinical trials:
- Porcine Model: In 11 porcine kidneys, the model's predicted ischemic zones were compared against actual "ground truth" zones created by ligating specific arteries. The model achieved a Dice coefficient of 0.82, demonstrating high spatial accuracy.
- Clinical Application: The model was prospectively applied to 131 human patients. Surgeons used the TKP maps to guide targeted clamping. Using Indocyanine Green (ICG) dye, they visually confirmed that the clamped areas matched the TKP model's predictions with a Dice coefficient of 0.81.
Clinical outcomes were highly favorable, with a mean operation time of 100.8 minutes and a median blood loss of only 110 ml.
Conclusion and Future Directions
The TKP model provides a feasible and effective tool for automated surgical planning in nephron-sparing surgery. By accurately predicting perfusion regions, it enables surgeons to minimize ischemic injury to healthy tissue. Future work will focus on multi-center trials and the integration of these 3D maps directly into surgical consoles using Augmented Reality (AR), providing real-time, "GPS-guided" navigation during complex procedures.
Summary updated