To study brain tumors from MRI images, oncologists must precisely delineate the contours of the lesions or "segment" them, i.e. assemble the pixels of the image into different groups, according to predefined criteria.

There are currently several deep learning segmentation architectures for brain tumors. These models only perform well for the tumor types upon which they have been trained. They are therefore better for common tumors, such as glioblastoma, than for rare tumors, such as infiltrating brainstem glioma (a pediatric cancer).

However, there are some visual similarities between common and rare tumors that make it possible to approach the problem in two steps: detection and then classification of pixels.

This is the approach adopted by the NeuroSpin researchers and their partners. They propose two delineation methods based on:

• object detection with an automatic object detection algorithm, known for its high accuracy and speed (You Only Look Once[);
• tumor segmentation with a convolutional neural network, developed for biomedical image processing (U-Net ).

For each step, the neural networks trained on common lesions were used on rare lesions, without any additional parameter adjustments. This strategy yielded better results when the tumor differed from the training tumor and robust delineations were obtained on the infiltrating brainstem glioma.

 
By addressing the problem of rare tumors, for which no database can be built to train a deep neural segmentation network, the researchers show that "simple" object detection and tumor segmentation can be combined to obtain good results without any retraining or adaptation of the model.