ImageCLEF is an international campaign organized by the CLEF initiative lab, aimed at developing novel solutions based on AI and machine learning techniques to solve several problems related to the healthcare and medical domain. This year’s ImageCLEFmed competition has the goal of developing the best tuberculosis type classification from CT scans.
SenticLab was also involved in the such a competition. Indeed, among the many research group that participated in this challenge of developing the best tuberculosis type classifier, our solution resulted to be the one with the highest accuracy score!
The competition involved a total of 59 different submissions, aimed at identifying the type of tuberculosis among five possible classes, a sample of those is depicted in the following image:
The task is particularly challenging, because some of the above classes show very few distinctive features, making the trained models particularly prone to confusion. Moreover, in many of the CTs made available within the dataset, there were several cases where the patient presented more than one affection, although only one label was chosen for each image.
The solution developed by SenticLab consists in a hierarchical approach where a first classifier was trained using only those CT scans with just one affection. This first model can detect if the CT presents any of the five aforementioned affections. If so, a second model (based on a separate training process) is used to detect the most important affection from the same CT, also aggregating the output of the first model.
This is not the first time our team wins the ImageCLEF competition. In 2020, when the competition goal was related to lung affections identifications within CT scans, the team won too, and results of that competition lead to a paper published on a book chapter published by Springer International.
We and our partner synbrAIn are seamlessly working together in the development of new AI-based solution for healthcare. Our collaboration has allowed both the teams to grow together, leading to the development of several machine learning approaches to process medical data and develop new predictive models to support image-based diagnosis.