Deep Learning for Automated Detection and Segmentation of Meningioma on Multiparametric MRI.
TL;DR
Deep learning models performance in automated detection and segmentation by using different combinations of routine MRI sequences suggests that, when optimized, multi-sequence models could improve the performance of single-sequence models, which may facilitate future development of automated workflows for volumetric meningioma assessment.
Abstract
Meningioma is the most common benign primary intracranial tumor. Although its typical appearance on contrast-enhanced T1-weighted images is characteristic, tumor conspicuity and boundary definition vary substantially across MRI sequences, which complicates automated detection and delineation. Automatic meningioma detection and segmentation provide the volumetric delineation on which tumor volume estimation, radiotherapy target definition, and radiomic analyses depend and may reduce the time and inter-observer variability associated with manual delineation. We investigated deep learning models performance in automated detection and segmentation by using different combinations of routine MRI sequences. We retrospectively analyzed 149 patients with histologically diagnosed meningioma. The imaging protocol included T1-weighted (T1), contrast-enhanced T1-weighted (T1-CE), T2-weighted (T2), T2 gradient echo-weighted (T2-GE), fluid-attenuated inversion recovery (FLAIR) sequences, and apparent diffusion coefficient (ADC). A maximum of 63 combinations were used to feed an Attention U-Net, and models' outputs were compared with manual segmentation by an experienced neuroradiologist to assess the detection rate (DR) and segmentation performance. Among single-sequence models, T1-CE (DR: 0.79 ± 0.11, DICE: 0.67 ± 0.10) showed the best performance. Models trained on multiple MRI sequences (up to 3) consistently improved those trained on single sequences. T1-CE + T1 + FLAIR model achieved best lesion detection (DR: 0.98 ± 0.03) and segmentation (DICE: 0.82 ± 0.04). By combining T1-CE with up to two additional sequences, the algorithm developed provides high efficiency and accuracy in automatic meningioma detection and segmentation. This work suggests that, when optimized, multi-sequence models could improve the performance of single-sequence models, which may facilitate future development of automated workflows for volumetric meningioma assessment.