Automatic detection of multi-site oral squamous cell carcinoma based on privacy-preserving federated learning.

BACKGROUND

Automated detection of oral squamous cell carcinoma (OSCC) using computer-aided technology plays a crucial role in ensuring patient health and reducing medical costs. However, traditional centralized disease recognition methods often face challenges such as data privacy, scalability, and substantial data transmission requirements. Federated learning offers a promising solution to these issues by enabling collaborative model training on distributed data sources.

METHODS

This paper proposes a privacy-preserving federated learning framework for multi-site automated detection of OSCC, analyzing a dataset of 1,224 oral images from the B. Borooah Cancer Institute. The framework leverages the distributed nature of data in medical environments, where multiple medical clinics capture oral images to identify diseases. The training process involves local computation on each device, rather than transmitting raw data to a centralized server, thus protecting data privacy and reducing communication overhead. The federated learning approach involves a central server that coordinates the training process across multiple edge devices. Initially, the central server distributes a pre-trained model to each device. Then, the devices perform local model updates using their own data, capturing unique disease patterns specific to their patient populations. These updated models are aggregated by the central server after adding Gaussian noise, combining the knowledge from all devices to create a global model representing the collective intelligence of the network.

RESULTS

Experimental results demonstrate the effectiveness of the proposed federated learning framework for automated OSCC detection. The distributed training method achieved comparable disease recognition accuracy to traditional centralized methods, with an area under the curve (AUC) of 0.887, accuracy (ACC) of 95.51%, sensitivity (SEN) of 95.92%, and specificity (SPE) of 95.10%, while effectively preserving data privacy.

CONCLUSION

The application of federated learning in automated OSCC detection opens up new possibilities for collaborative and privacy-preserving machine learning solutions. This approach enables medical institutions to leverage the collective knowledge of distributed devices while maintaining data confidentiality, thereby promoting efficient and accurate disease recognition in medical settings.