Medical image pretraining-based transfer learning for generalizable and robust diagnosis of bone tumors on radiographs: a multi-center study

Key Points

• Evaluate the efficacy of medical image pretraining for bone tumor classification on radiographs.
• Utilized RadImageNet for pretraining models on medical images.
• Compared performance with models pretrained on natural images (ImageNet).
• Assessed the effectiveness of AI assistance across varying radiologist experience levels.
• Medical image pretraining outperformed natural image pretraining for tumor classification.
• AI support effectiveness showed variability among radiologists based on their experience.
• Deep learning models demonstrated higher accuracy compared to traditional radiomics approaches.

Abstract