Federated Learning for Renal Tumor Segmentation and Classification on Multi-Center MRI Dataset.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Oncology research

Background:

• Deep learning (DL) models require multi-center data for generalizability in renal tumor characterization.
• Data-sharing limitations necessitate privacy-preserving techniques like federated learning (FL).

Purpose of the Study:

• To evaluate the performance and reliability of FL for renal tumor segmentation and classification.
• To compare FL with non-FL approaches on multi-institutional MRI datasets.

Main Methods:

• A retrospective multi-center study included 987 patients with renal tumors.
• FL and non-FL models (nnU-Net for segmentation, ResNet for classification) were trained and tested.
• Performance was assessed using Dice coefficients for segmentation and AUC, accuracy, sensitivity, and specificity for classification.

Main Results:

• No significant differences were found between FL and non-FL models in segmentation (Dice: 0.43 vs. 0.45, p=0.202).
• Classification performance also showed no significant difference (AUC: 0.69 vs. 0.64, p=0.959).
• Accuracy, sensitivity, and specificity metrics were comparable between FL and non-FL approaches.

Conclusions:

• Federated learning (FL) offers comparable performance to traditional centralized training for renal tumor segmentation and classification.
• FL is a promising privacy-preserving solution for multi-institutional deep learning model development.
• This validates FL's potential in enhancing the generalizability of AI models in medical imaging.