Tribun Health offers hybrid or full cloud solutions hosted by Microsoft Azure which can provide scalable infrastructure, cost-effective storage and powerful computational resources Read More
Tribun Health establishes itself as a confirmed AI actor
One year ago, Tribun Health set up a dedicated AI team composed of data scientists and engineers specialized in computer vision
Our team is working on the development of image analysis algorithms on virtual slides to respond to various medical issues such as scoring of prognostic and predictive biomarkers of breast cancer or tissue segmentation and sorting of colonic biopsy slides. It is in this context that we naturally participated in the TissueNet 2020 Data Challenge on cervical lesion detection. The work was conducted on data that was available during the Data Challenge 2020 organized by the French Society of Pathology (SFP) and the Health Data Hub (HDH), with the support of the Grand Defi for AI in Health.
We are very proud to have won this competition by developing a quality deep learning algorithm with a score of 0.947 for the classification of cervical slides. This international challenge gathered more than 500 participants from all over the world. It was an excellent opportunity for our team to demonstrate their skills and experience on a stimulating set of data.
Our algorithmic approach, which we illustrated in our poster entitled “Detect Lesions in Cervical Biopsies“, consists in working in a supervised learning framework in which we used only the 1000 annotated slides of the data challenge without any additional annotation. Our strategy was based on working closely with pathologists and understanding the pathological tissues and grade differences with them. Preliminary information on the location of the grades as well as the right balance between the context of the lesion and the image resolution were efficiently modeled by our data scientists. In addition, an efficient analysis pipeline based on a predefined DenseNet model has also been key to our success and has enabled us to achieve a promising score which is an encouraging step towards the clinical use of our algorithms to save time for pathologists.