3753 - Deep Transfer Learning and Dosiomics for Predicting Radiation-Induced Temporal Lobe Injury in Nasopharyngeal Carcinoma

L. Wang¹, J. Zhou², T. Qiu¹, Y. Zhu¹, B. Sun³, G. Yang⁴, S. Huang⁵, X. He¹, and L. Wu⁵; ¹The Affiliated Cancer Hospital of Nanjing Medical University, Nanjing, China, ²Department of Radiation Oncology, The Affiliated Cancer Hospital of Nanjing Medical University & Jiangsu Cancer Hospital & Jiangsu Institute of Cancer Research, Nanjing, China, ³Department of Radiation Oncology, Dan L. Duncan Comprehensive Cancer Center, Baylor College of Medicine, Houston, TX, ⁴Southeast University, Nanjing, China, ⁵Department of Radiation Oncology, Jiangsu Cancer Hospital & Jiangsu Institute of Cancer Research & Nanjing Medical University Affiliated Cancer Hospital, Nanjing, China

Purpose/Objective(s): To investigate the potential of ADC map-based deep transfer learning and dose distribution-based dosiomics in predicting radiation-induced temporal lobe injury (RTLI) in nasopharyngeal carcinoma (NPC) .

Materials/

Methods: A total of 194 NPC patients were enrolled, including 97 patients with RTLI and 97 patients without RTLI. Patients were randomly assigned to the training cohort (n = 135) and the validation cohort (n = 59). Deep transfer learning (DTL) features and dosiomics features were extracted from ADC map and dose distribution, respectively. We utilized the Spearman rank correlation test and least absolute shrinkage and selection operator regression for feature selection. Subsequently, a total of eight machine learning classification models were trained to establish a prediction framework, encompassing Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF), Extremely Randomized Trees (Extra Trees), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost) and Multilayer Perceptron (MLP).The performance of clinical, deep transfer learning, dosiomics and feature fusion model was compared by the area under the curve (AUC).

Results: We constructed eight pre-trained transfer learning networks and extracted DTL features, respectively. The results showed that pre-trained WideResNet 101 exhibited superior performance with an AUC of 0.786 (95% CI: 0.666 - 0.905) in the validation cohort. The clinical model based on D1cc and induction chemotherapy demonstrated an AUC of 0.794 (95% CI: 0.681 - 0.907) and the dosiomics model demonstratedan AUC of 0.903(95% CI: 0.827 - 0.979). Features fusion model demonstrated the highest AUC values in both the training (0.988, 95% CI: 0.970 - 1.000) and validation (0.940, 95% CI: 0.879 - 1.000) cohorts.

Conclusion: We first utilized comprehensive multidimensional data based on deep transfer learning, dosiomics, and clinical features with a good predictive ability for predicting the RTLI in NPC patients, which can support clinician decision-making in developing individualized treatment plans and providing preventive measures.