In neuro-oncology, tumors are often investigated by magnetic resonance imaging (MRI). The segmentation of tumor lesions is a crucial step for the study of tumors from images. It consists of a precise delineation of the tumor region in a diagnostic or surveillance image. Currently, several deep learning segmentation architectures are proposed for automatic brain tumor segmentation. These models perform well for the type of tumor they are trained on, such as common tumors like multiform glioblastoma. However, their performances are not guaranteed when used with rare tumors, such as brainstem infiltrating glioma, for which the number of labeled examples is insufficient for de novo training or to adapt the parameters of a pre-trained model.

Because some visual similarities nevertheless exist between common and rare tumors, one can approach the problem in two steps: detection and then classification of pixels. The present study exploits some similarities of the two types of tumors, observable at multiple scales, and proposes two delineation methods (pYU and sYBBU) based on the combination of object detection with the YOLO network and tumor segmentation with a U-Net network. For each step, networks trained on common lesions were used on rare lesions, following a domain adaptation scheme without additional parameter adjustments. This strategy yielded better results when the tumor differed from the training tumor and robust delineations were obtained on brainstem infiltrating glioma, a rare pediatric tumor located in the brainstem region.

Schematic of the two approaches pYU and sYBBU used: parallel model YOLO UNet (pYU), orange box; sequential model YOLO BB-Unet (sYBBU), red box. The authors chose the well-known YOLO object detection model in its pre-trained version on natural images. Unet and BB-Unet are U-shaped networks trained on common tumor images. The combination of these two types of networks allows us to obtain robust segmentations of rare tumors. H Chegraoui et al, Cancers, 2021

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

Contacts : hamza.chegraoui@cea.fr ; vincent.frouin@cea.fr

 - Image segmentation consists in grouping pixels according to predefined criteria in order to build regions or classes of pixels.  

 - The localization of one or more objects in the images was performed with YOLO (You Only Look Once), an automatic object detection algorithm known for its high accuracy and speed (J.Redmon et al., 2015).

 - U-Net is a convolutional neural network developed for biomedical image segmentation.