In this project, we aim to explore the opportunities in applying machine learning techniques in medical context, especially in digital pathology. We developed a automated system to diagnose Bacterial Vaginosis infection, a common infection found in women at reproductive age, by examining the smears.
We start by calibrating the variations between microscopic images in the training dataset. This could include using Gaussian filters and granulometric analysis.
The normalised images are then segmented using differenet methods and a variety of blob detection algorithms including sliding window, DBSCAN and MSER, into interest areas which typically contains a patch of the image or a single bacterium.
In the first two phases of the project, a classifier which distinguishes between target bacteria types from the rest is trained. In the final phase of the project, a object detector is trained, predicting both the bounding boxes of the bacteria and the types of them. A number of different machine learning algorithms and models are used, including convolutional neural networks, residual networks and faster R-CNN.
The classified images are then collected, and the overall interpretation in the degree of infection is determined using the Nugent Score system.
Clicklable is a highly reusable data labelling tool for image classification and object detection purposes.
Users can choose different labelling modes including "Point" labelling or "Rectangle" labelling which facilitates both interest-area classification and object detection.
This tool is platform-independent, where any computer with Java™ Runtime Environment installed can run this tool with the most native interface.
This diagnostic system is a all-in-one system for disgnosing Bacterial Vaginosis.
Users can load a trained machine learning model and starts the diagnosis by picking a smear image. Detailed results are drawn on the image and interpretations are also given. Users can also export the image for other purposes.
This tool is also platform-independent, where any computer with Python and the corresponding dependencies installed can run this tool with the most native interface.
Chi Ian Tang (Michael) is currently in his final year at the Department of Computer Science, the University of Hong Kong.