A Machine Learning Model to Predict the Histology of Retroperitoneal Lymph Node Dissection Specimens

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Post-chemotherapy retroperitoneal lymph node dissection (PC-RPLND) is crucial for teratoma or viable germ cell tumors (GCT).
• Overtreatment occurs if PC-RPLND reveals only necrosis.
• Predicting residual teratoma in PC-RPLND specimens is challenging.

Purpose of the Study:

• Develop machine learning models to predict residual teratoma in PC-RPLND specimens.
• Utilize CT imaging and clinical variables for prediction.
• Improve diagnostic accuracy for residual teratoma in GCT patients.

Main Methods:

• 58 patients undergoing PC-RPLND were analyzed.
• ResNet50 (CT imaging) and Support Vector Machine (SVM) (clinical variables) algorithms were applied.
• A nested, 3-fold cross-validation protocol was used.

Main Results:

• ResNet50 achieved 80.0% diagnostic accuracy (67.3% sensitivity, 90.5% specificity, 0.84 AUC).
• SVM achieved 74.8% diagnostic accuracy (59.0% sensitivity, 88.1% specificity, 0.84 AUC).
• Histology confirmed teratoma in 71/155 regions of interest.

Conclusions:

• Machine learning models reliably distinguish residual teratoma from necrosis in PC-RPLND specimens.
• AI-driven analysis of CT imaging and clinical data enhances diagnostic precision.
• These models can potentially reduce unnecessary overtreatment in GCT management.