Object Detection Improves Tumour Segmentation in MR Images of Rare Brain Tumours

SIMPLE SUMMARY: This study evaluates the impact of adding an object detection framework into brain tumour segmentation models, especially when the models are applied to different domains. In recent years, multiple models have been successfully applied to brain tumour segmentation tasks. However, the performance and stability of these models have never been evaluated when the training and target domain differ. In this study, we identify object detection as a simpler problem that can be injected into a segmentation model as an a priori, and which can increase the performance of our models. We propose an automatic segmentation model that, without model retraining or adaptation, showed good results when applied to a rare brain tumour. ABSTRACT: Tumour lesion segmentation is a key step to study and characterise cancer from MR neuroradiological images. Presently, numerous deep learning segmentation architectures have been shown to perform well on the specific tumour type they are trained on (e.g., glioblastoma in brain hemispheres). However, a high performing network heavily trained on a given tumour type may perform poorly on a rare tumour type for which no labelled cases allows training or transfer learning. Yet, because some visual similarities exist nevertheless between common and rare tumours, in the lesion and around it, one may split the problem into two steps: object detection and segmentation. For each step, trained networks on common lesions could be used on rare ones following a domain adaptation scheme without extra fine-tuning. This work proposes a resilient tumour lesion delineation strategy, based on the combination of established elementary networks that achieve detection and segmentation. Our strategy allowed us to achieve robust segmentation inference on a rare tumour located in an unseen tumour context region during training. As an example of a rare tumour, Diffuse Intrinsic Pontine Glioma (DIPG), we achieve an average dice score of 0.62 without further training or network architecture adaptation.