The CT-based deep learning model outperforms traditional anatomical classification models in preoperatively...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Urology and Nephrology
• Oncology

Background:

• Partial nephrectomy is a standard treatment for renal tumors.
• Predicting perioperative complications and risk stratification is crucial for patient management.
• Traditional anatomical classification models have limitations in accuracy.

Purpose of the Study:

• To develop and validate a deep learning (DL) model integrating CT radiomics and clinical features for predicting perioperative complications and risk grade in partial nephrectomy.
• To compare the DL model's performance against traditional RENAL and PADUA scoring systems.

Main Methods:

• A DL model was developed using CT radiomics (nnU-Net, pyradiomics) and clinical data from 1214 patients.
• Logistic regression models based on RENAL and PADUA scores were created for comparison.
• External validation was performed on a cohort of 260 patients to assess generalizability.

Main Results:

• The DL model achieved an AUC of 0.87 for predicting complications, significantly outperforming RENAL (0.68) and PADUA (0.69) models (p < 0.001).
• For risk grading, the DL model showed superior performance in the no-risk (AUC 0.83) and low-risk (AUC 0.79) groups compared to RENAL/PADUA.
• No significant differences were observed for moderate- and high-risk groups. External validation yielded an AUC of 0.83.

Conclusions:

• The CT-based DL model demonstrates superior performance in predicting complications and risk grades for no-risk and low-risk patients undergoing partial nephrectomy.
• The model offers a promising tool for improved risk stratification and perioperative planning in partial nephrectomy.