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

Abstract

Background/aim: While post-chemotherapy retroperitoneal lymph node dissection (PC-RPLND) benefits patients with teratoma or viable germ cell tumors (GCT), it becomes overtreatment if necrosis is detected in PC-RPLND specimens. Serum microRNA-371a-3p correctly predicts residual viable GCT with 100% sensitivity; however, prediction of residual teratoma in PC-RPLND specimens using current modalities remains difficult. Therefore, we developed a machine learning model using CT imaging and clinical variables to predict the presence of residual teratoma in PC-RPLND specimens.

Patients and methods: This study included 58 patients who underwent PC-RPLND between 2005 and 2019 at the University of Tsukuba Hospital. On CT imaging, 155 lymph nodes were identified as regions of interest (ROIs). The ResNet50 algorithm and/or Support Vector Machine (SVM) classification were applied and a nested, 3-fold cross-validation protocol was used to determine classifier accuracy.

Results: PC-RPLND specimen analysis revealed 35 patients with necrosis and 23 patients with residual teratoma, while histology of 155 total ROIs showed necrosis in 84 ROIs and teratoma in 71 ROIs. The ResNet50 algorithm, using CT imaging, achieved a diagnostic accuracy of 80.0%, corresponding to a sensitivity of 67.3%, a specificity of 90.5%, and an AUC of 0.84, whereas SVM classification using clinical variables achieved a diagnostic accuracy of 74.8%, corresponding to a sensitivity of 59.0%, a specificity of 88.1%, and an AUC of 0.84.

Conclusion: Our machine learning models reliably distinguish between necrosis and residual teratoma in clinical PC-RPLND specimens.

Keywords: Germ cell tumor; Resnet50 algorithm; machine learning model; post-chemotherapy retroperitoneal lymph node dissection; support vector machine.