Postoperative glioma project
Our client – one of the leading companies in the pharmaceutical industry – approached us with a need to create a model from scratch that would segment areas of a patient’s brain, following surgery related to a brain glioma resection. The algorithm had to not only segment the brain lesions, but also provide objective volume data. A collection of brain MRI studies had been provided by the client.
What we did
MRI brain imaging studies after contrast enhancement – after resection of a brain glioma – were unlabeled. We implemented and carried out the process of labeling them by experts (radiologic physicians). It was on them that we trained the model we had developed. The model was created using advanced machine learning techniques, deep learning in particular. It was intended that the algorithm we created would segment three areas: the boundaries of the region that resulted from tumor excision (cavity), edema, and necrosis. During surgery, a situation may occur in which not the entire tumor is resected and parts of it are still visible on the medical image. Here, a challenge for the algorithm arose, in the form of the need to distinguish the cavity border from the tumor fragments not removed during surgery. Both tumor remains and borders are naturally enhanced with contrast. In the process of training the algorithm, this effect was achieved.
Our work resulted in a precise algorithm that not only segments the three identified areas, but also provides objective data in the form of the calculated volume of each area.