3682 - Compare the Predictive Performance of Deep Learning and Radiomics Models for Predicting Radiation-Induced Temporal Lobe Injury in Nasopharyngeal Carcinoma

J. Liu¹, 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 compare the predictive performance of deep learning, deep transfer learning and radiomics models based on apparent diffusion coefficient (ADC) maps to predict radiation-induced temporal lobe injury (RTLI) in patients with nasopharyngeal carcinoma (NPC). Materials/

Methods: We retrospectively included 230 NPC patients,115 with RTLI and 115 without RTLI. Patients are randomly divided into the training cohort (n=161) and the validation cohort (n=69) in a ratio of 7:3. Deep learning, deep transfer learning and radiomics features were extracted from ADC maps. The Pearson correlation coefficient and the least absolute shrinkage and selection operator (LASSO) regression were used to select features. Subsequently, we trained eight machine learning classification models, including Logistic Regression, Support Vector Machine, K-Nearest Neighbor, Random Forest, Extremely Randomized Trees, eXtreme Gradient Boosting, Light Gradient Boosting Machine and Gradient Boosting. The area under the curve (AUC) was calculated to compare the predictive performance of various models.

Results: MobileNet V3 and pre-trained MobileNet V3 showed the best predictive performance in deep learning networks and deep transfer learning networks, respectively, and they demonstrated the AUC of 0.909 (95% CI: 0.841 - 0.977) and 0.933 (95% CI: 0.867 - 0.999) in the validation cohort, respectively. The radiomics model demonstrated an AUC of 0.831 (95% CI: 0.733 - 0.929). Therefore, the deep transfer learning model achieved the highest AUC values in both the training (0.918, 95% CI: 0.875 – 0.961) and validation (0.933, 95% CI: 0.867 - 0.999) cohorts. The ClinicalTrial.gov ID of this study is NCT06244394.

Conclusion: The potential of deep learning in predicting RTLI surpasses that of radiomics, and the application of transfer learning can further enhance the predictive capability.