Deep learning segmentation of paediatric brain tumours using diffusion-weighted MRI: towards an early, in vivo classification pipeline.
This retrospective study used transfer learning and multimodal ensembling with a 3D DeepMedic network to segment paediatric brain tumours from diffusion-weighted MRI, achieving a best-model median Dice score of 0.63 against manual annotations in 107 patients.
Open original publication →What the AI sees
This retrospective study used transfer learning and multimodal ensembling with a 3D DeepMedic network to segment paediatric brain tumours from diffusion-weighted MRI, achieving a best-model median Dice score of 0.63 against manual annotations in 107 patients.
Research significance
The evidence supports DWI-only automated tumour segmentation as a possible enabling component for quantitative imaging workflows; it may eventually improve early classification or clinical decision support, but treatment selection, outcome improvement, and clinical adequacy were not tested.
Source abstract
Diffusion-weighted MRI (DWI) can offer vital quantitative biomarkers to understand paediatric brain tumours to benefit clinical management. However, extraction of these markers requires delineation of the tumour margins, which is time-consuming, requires expertise and can still be highly variable. Automated approaches using deep learning to generate these tumour regions exist, but to our knowledge, few exist using only DWI as an input modality, with none in paediatrics. This retrospective study develops an automated segmentation approach, leveraging transfer learning and multi-modal ensembling to tackle the issues specific to this DWI-only approach. Using the Imaging of Tumors study data, we analysed data from 107 paediatric brain tumour patients. Using a 3D convolutional neural network, namely DeepMedic, to perform automatic segmentations, we demonstrate the benefit of these approaches for this task. We assess the accuracy of these predicted segmentations in comparison to the 'ground truth' manual annotations, with a median Dice score of 0.63 achieved by the best-performing model. The current study highlights the potential of this approach to be implemented in future clinical decision support tools using DWI, but more work is needed to improve segmentation accuracy or establish current performance as 'sufficient' for the purposes that these segmentations are required.