Privacy-preserving polyp segmentation using federated learning with differential privacy

Abstract

Background and Objective:

Patient privacy is of paramount importance in the medical field, especially as data-driven medical applications gain popularity. The privacy of medical records is increasingly crucial. In this context, data-oriented polyp (a precancerous stage of colon cancer) segmentation is a critical area of ongoing research, aiming to improve automated segmentation. Accurate segmentation is essential for the complete removal of these overgrown cells from the gastrointestinal system. Although large data sets using data-driven algorithms have shown excellent performance in image segmentation, privacy concerns have limited the availability of such datasets for medical image segmentation tasks, including polyp segmentation. This research aims to develop an approach for polyp segmentation that combines data from multiple sources without compromising patient privacy.

Methods:

We design a differentially private federated learning system to segment polyps without compromising privacy. Our approach employs the encoder–decoder architecture UNet 3+ with a deep supervision technique to achieve effective segmentation of polyps in a federated setup. The federated training process aims to find generalized global models for the entities participating in the federation. The study uses four public databases to train and evaluate the proposed method.

Results:

The proposed privacy-protected technique demonstrates promising outcomes in polyp segmentation, achieving an average Intersection over Union (IoU) score of 0.90881 ± 0.00355 over four publicly available datasets. Evaluation metrics include precision, sensitivity, and specificity values, indicating the effectiveness of our approach in accurately segmenting polyps.

Conclusions:

Our differentially private federated learning system successfully segments polyps without compromising patient privacy. The promising results suggest that this approach can significantly contribute to the field of polyp segmentation, facilitating the use of large datasets while maintaining strict privacy standards.