Radiologists’ nightmare — segmentation masks
This blog is the second in our xLungs series about the Responsible Artificial Intelligence for Lung Diseases project. You can check out the first one about our way towards the largest Polish database of lung medical images here.
In this blog, we focus on creating ground truth segmentation masks on the collected data. This is another step towards building the AI-based solution to help radiologists in their daily work. There are few convolutional network architectures for lung segmentation however none of them is perfect. The main problem is the fact that they “see” only healthy lung parenchyma. Large subpleural consolidations, tumors, and peripheral nodules are overlooked.
Therefore, in our xLungs — Responsible Artificial Intelligence for Lung Diseases project that is carried out at MI2DataLab at Warsaw University of Technology, with help of radiologists, we want to create a more precise method to perform automatic lung segmentation. That would be impossible without high-quality data input which is in line with a shift of the Model-centered AI paradigm to Data-centered AI.
A significant part of this project is to create lung segmentation in Computer Tomography (CT) examinations that include abnormalities in lung parenchyma and pleural cavity. At this stage 55 CTs were examined. Each of them consists of about 300–350 slices that has to be analysed by radiologists in three planes, in regards to anatomic structures (right and left lung and background), main pathologies within lungs (i.e. atelectasis) and pleural cavity (i.e. pleural effusion, pneumothorax). This gives about 4000 slice-plane-structure fragments of CTs that have to carefully investigated. In total, radiologists examined about 220 000 fragments of CTs and made about 12 500 pixel-level annotations in all 55 studies.
That process is engaging and time-consuming. Every segmentation mask is created using open-source 3D-slicer software. Step by step every significant area is manually marked and checked. All studies are processed in the same way. We hope that our accurate segmentation will help to train a responsible model for automatic segmentation.
Credits to our radiologists: Przemysław Bombiński MD PhD, Patryk Szatkowski MD.
If you are interested in other posts about explainable, fair, and responsible ML, follow #ResponsibleML on Medium.
