Abstract Reliable tumour detection is central to clinical decision-making, therapy planning, and monitoring of disease progression. It remains a persistent challenge due to widely variable tumour morphology, size, and contrast, and the often-subtle appearance of lesions in 2D/3D magnetic resonance imaging (MRI). Strategies that employ artificial intelligence are being sought to improve tumour targeting and delineation, small lesion detection as well as a robust analysis framework that generalise across adult and paediatric brain tumour cases. Variability in technology leads to differences in image quality, whilst incomplete MRI sequences, irregular pixel/voxel-wise annotations and limited number of images, specifically from paediatric brain tumour patients, confound current attempts to apply state-of-the art supervised AI model to real-world clinical paradigms. We address such limitations through advancing a methodology for ‘weakly-supervised’ brain tumour segmentation using denoising diffusion probabilistic models. Reducing the dependence on fully supervised, densely annotated datasets, we unlock the potential for the analysis to learn from sparse, heterogeneous datasets whilst improving robustness and generalisability of the outcome. We present our strategy for testing a bespoke weakly-supervised anomaly detection method which offers promising alternatives to fully supervised segmentation by reducing reliance on radiologist derived voxel-level annotations whilst enabling robust localisation of abnormal tissue. We explain our strategy to implement super-resolution image segmentation to accurately detect smaller brain tumour lesions, typical in paediatrics, that demonstrated clear gains in sensitivity and resolution-aware segmentation of all lesions. Finally, we conduct experiments on the ‘Brain Tumour Segmentation’ dataset to assess the generalisability of this framework to paediatric tumours, an underexplored domain limited by sparce datasets. Experimental results show our approach, when trained solely on adult data, generalise effectively to paediatric cases. Additional evaluation on a multi-institutional cohort encompassing diverse tumour types further supports the framework’s robustness but confirms the need for hospital datasets that build a ground truth.
Catchpoole et al. (Tue,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: