Automated Sacral Tumor Localization and Multi-Class Identification of Six Tumor Types in NCCT with an innovative CL-MedImageNet Fusion Model

Motivation: Accurate preoperative identification of tumor types is essential for personalized treatment. Goal(s): To develop a fully automated hybrid approach to predict sacral tumor types from preoperative NCCT images. Approach: We built an automated hybrid model integrating two deep CNN models (Model 1 and Model 2) through a fully automated pipeline. Results: The CL-MedImageNet model achieved the best results, with macro average AUC of 0.891, 0.883, and 0.874 for the validation set, internal test set, and external test set, respectively. The CL-MedImageNet model also surpassed the radiologist's performance. Impact: Our model reliably predicts sacral tumor types using a fully automated NCCT process, improving individualized treatment planning with its high reproducibility and generalizability.