All news

Neural Net­works for Neuro­path­o­logy

28.07.2020, Re­search :

With the aid of neural networks, malignant brain tumors can be quickly detected and automatically marked for pathological evaluation.

The project "Neural Networks for Neuropathology" (N4N) is carried out by HNU’s DigiHealth Institute in collaboration with the Institute of Neuropathology of the Technical University of Munich (TUM).

The goal of the project is to investigate pathological sections of glioblastomas – the most common malignant brain tumor with an infaust prognosis1 – using modern machine learning methods. Tissue samples taken intraoperatively must be evaluated as quickly as possible in order to provide the surgeon with feedback on the degree of resection. The operation can be individually adapted and supported by the new method. A neural network will automatically pre-evaluate optical microscopic data, highlight tumor-affected tissue areas, and direct the pathologist's attention to suspicious areas within the pathological sections. This procedure will enable a more precise and faster diagnosis. In a second step, the image data and other metadata can be used to predict patient survival.

The Institute of Neuropathology at TUM provides the image data for this purpose. Individual image scenes of the pathological slices are about 6.5 gigabytes in size and have a resolution of 70,000 by 40,000 pixels. In order to make the image data effectively usable, the images are first labelled, after which the HNU trains a neural network for the automated detection of malignant tumor cells. The results are then available to the pathologist for evaluation.

 


1Glioblastoma multiforme (IDH wt.) (lat. for multiform) is the most common malignant brain tumor (WHO grade IV) in adults. The median survival period is 15 months.

 

 (opens enlarged image)
Non-labelled (left) and manually labeled (right, black) tumor areas within a pathological section. Image width is approximately 1 mm. Image: Georg Prokop, Neuropathology Technical University Munich.