A distributed deep learning approach for histopathology image retrieval

Abstract

Digital microscopy and radiology generate growing amounts of imagery data. To help practitioners find the information crucial to establish the most accurate possible diagnoses, Artificial Intelligence tools need to be developed.

This master thesis, based on the study of existing literature and open-source code, proposes a distributed deep learning architecture that allows a user, by using a fast approximate nearest neighbour search, to retrieve similar histopathology images to a query image.

The retained Deep Learning architecture, ResNet50 with some modifications, was distributed on different servers in order to allow the handling of up to million or billion images.

It was trained on a large-scale dataset of 67 classes of annotated medical images and the obtained results are quite promising, as well for the visual similarity of the retrieved images as for the search time. This research also analyses the generalisation to classes on which the system was not trained, and the impact of the approximated search on the accuracy and the retrieval time.

Nevertheless, even though the results are positive, this system might present some limitations as it was tested on only one dataset and was not reviewed by medical practitioners.