Developing a scalable and privacy-preserving deep learning model for the classification of peripheral blood cell images

Abstract

Histopathological diagnosis is a time-intensive process dependent on the expertise and interpretative criteria of pathologists. Digital pathology, employing machine learning models, offers a promising avenue to enhance diagnostic accuracy and efficiency through computer-aided diagnosis systems. Specifically, the automated counting and identification of cells from blood smears constitute 80% of the initial analyses required for detecting haematological diseases.

The intrinsic sensitivity of medical data demands robust privacy safeguards. This has focused recent investigations into the potential of collaborative learning, or Federated Learning (FL), as a scalable and inherently private training paradigm. By training data locally and subsequently aggregating parameters on a central server, the direct movement and sharing of medical data are circumvented. Nevertheless, recent studies have cast doubt on the privacy of these collaborative trainings. Moreover, collaborative learning faces the challenge of dealing with the heterogeneity of participating clients to generate an efficient model across various nodes.

This work presents a performance comparison between different training types for peripheral blood cell classification models. The findings suggest that collaborative learning, both in homogeneous (IID) and heterogeneous (non-IID) clients, could enhance the predictive capability of conventionally trained classification models. Furthermore, collaborative learning has the potential to reduce the time and resources required for model training.