Smart Surveillance of Secondary Brain Tumours

Segmentation of secondary brain tumours enables clinicians to plan clinical interventions. This task is currently done manually, which is time consuming and susceptible to inter-observer variability. Automating the segmentation is challenging due to the varying size, number, and heterogeneity of the lesions. Presently, 1 W post-Gadolinium images are used by clinicians for manual segmentation. This approach exposes the patient to gadolinium (Gd), which has been found to accumulate in healthy tissue and could potentially have long term health effects. To avoid this and address the need for a reliable, fast, and reproducible segmentation approach, we present a deep-learning-based automatic segmentation algorithm from magnetization transfer ratio (MTR) images, Gd-enhanced T1-weighted images (T1w-Gd), and T2-weighted Fluid-Attenuated Inversion Recovery (FLAIR) images. A comparison between the various combinations of imaging modalities indicate that automatic tumour, and tumour sub-region segmentation is doable with deep learning, yet the accuracy is limited, likely due to the small dataset and the heterogeneity of the tumours. Additionally, the use of gadolinium as an exogenous contrast agent remains necessary while automatic segmentation using MTR continues to be investigational.