Multi-reader evaluation of deep learning-based auto-segmentation of eloquent brain arteriovenous malformation on MRA and white matter tractography in stereotactic radiosurgery.
Authors
Affiliations (11)
Affiliations (11)
- Medical AI Lab, The First Hospital of Hebei Medical University, Hebei Medical University, Shijiazhuang, China.
- Hebei Provincial Engineering Research Center for AI-Based Cancer Treatment Decision-Making, The First Hospital of Hebei Medical University, Hebei Medical University, Shijiazhuang, China.
- Radiation Oncology Center, Huashan Hospital, Fudan University, Shanghai, China.
- Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China.
- Neurosurgical Institute of Fudan University, Shanghai, China.
- Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China. [email protected].
- Neurosurgical Institute of Fudan University, Shanghai, China. [email protected].
- Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China. [email protected].
- Neurosurgical Institute of Fudan University, Shanghai, China. [email protected].
- Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China. [email protected].
- Neurosurgical Institute of Fudan University, Shanghai, China. [email protected].
Abstract
To minimize the radiation injury for white matter (WM) pathways during brain arteriovenous malformation (bAVM) stereotactic radiosurgery (SRS), the WM tractography is integrated into treatment planning to identify WM pathways and restrict receiving dose. Manual segmentation of eloquent bAVM adjacent to WM pathways is time-consuming and prone to substantial inter-practitioner variability due to intricate entanglement within eloquent brain areas. The objective of this study is to develop and evaluate a deep learning (DL) system for the segmentation of eloquent bAVM in a clinical setting. A total of 191 eloquent bAVM patients who underwent WM tractography and 3D time-of-flight magnetic resonance angiography (TOF-MRA) images were enrolled. 153 patients were used to construct a two-stage DL bAVM segmentation ensemble (TBASE) consisting of 2D detection and 3D segmentation models to segment the bAVM, the other 38 to test performance. Comparative experiments with ResNet and U-Net were conducted to validate the effectiveness of the proposed network. A randomized multi-reader evaluation was then conducted to assess the impact of TBASE assistance for bAVM segmentation using ten algorithm-unseen cases. Six medical professionals contoured the same series of cases in both assisted and unassisted modes, with a 6-week memory washout period between each session. The aided and unaided Dice Similarity Coefficients (DSC), Hausdorff Distance (HD), along with contouring times were compared. The mean values and standard deviations for DSC and HD of TBASE are 0.87 ± 0.03 and 3.51 ± 0.26, respectively, while Res-Net and U-Net results are 0.75 ± 0.12 and 4.14 ± 0.99, 0.77 ± 0.09 and 3.94 ± 0.82, respectively. The average volume difference across all patients in test dataset is 0.25 ± 1.39 cc, with no statistically significant variation observed. With TBASE assistance, the mean DSC of readers improved from 0.76 ± 0.07 to 0.86 ± 0.05 (P < 0.001), with corresponding values of mean HD reducing from 4.31 ± 0.68 to 3.35 ± 0.17 (P < 0.001) and a mean time saving of 52.15% ± 13.85% per patient. Less-experienced readers achieved greater improvements in contouring accuracy compared to specialists (DSC increase: 0.15 ± 0.11 vs. 0.06 ± 0.09; P < 0.001), while demonstrating similar reductions in contouring time as specialists (50.68% ± 14.62% vs. 55.1% ± 11.98%; P = 0.217). The reliable eloquent bAVM automated-segmentation method has been validated in clinical workflow. The TBASE assistance improved the accuracy and efficiency of the eloquent bAVM manual delineation while considering WM pathway protection. Not applicable.