Ruffle, James; Mohinta, Samia; Gray, Robert; Hyare, Harpreet; Nachev, Parashkev; (2022) Brain tumour segmentation with incomplete data. Presented at: BRAIN Conference 2022, Online conference.

Preview

Text BrainConferenceJKR_slidedeck.pdf - Accepted Version Download (4MB) | Preview

Abstract

Brain tumour segmentation remains a challenging task, complicated by the marked heterogeneity of imaging appearances and their distribution across multiple modalities: FLAIR, T1-weighted, T2-weighted, and contrast-enhanced T1-weighted sequences (T1CE). This has compelled a research focus on uniformly multimodal models trained on complete acquisition sets rare in real-world clinical practice. Consider, for example, patients with renal failure who cannot receive contrast, artefact-spoiled sequences, or patients undergoing single-sequence intraoperative imaging. How well do segmentation models perform with such incomplete data, and what features of the lesion are identifiable under these circumstances? In a large-scale analysis involving 30 distinct segmentation models, we answer these questions with a state-of-the-art tumour segmentation modelling ensemble, nnU-Net-derived (Isensee et al, Nature Methods, 2020), deployed across all possible combinations of imaging modalities, trained, and tested with five-fold cross-validation on the 2021 BraTS-RSNA glioma population of 1251 patients. Segmentation performances for whole lesions range from Dice scores of 0.907 (single sequence) to 0.945 (full datasets) (Figure 1). When segmenting lesions by tissue type (enhancing tumour, non-enhancing tumour and oedema), Dice scores range from 0.701 (single sequence) to 0.891 (full datasets). Models missing postcontrast imaging still achieve a Dice coefficient for the whole tumour of 0.942 and identify the enhancing tumour component with Dice of up to 0.790 (Figure 2). Segmentation models can identify tumours with missing data, and can be used in clinical situations where partial data is frequent.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item