https://atlas.rsna.org/schemas/2025-11/model.json

Expert-centered Evaluation of Deep Learning Algorithms for Brain Tumor Segmentation

https://dx.doi.org/10.1148/ryai.220231

1.0

Jayashree Kalpathy-Cramer (corresponding author), email: ude.ztuhcsnauc@remarc-yhtaplak.eerhsayaj

NIH U01CA242879 (K.V.H., J.K.C.); NIH R01CA129371 and K23CA169021 (E.R.G.); P41EB015896 (Athinoula A. Martinos Center resources). Additional disclosures and institutional/grant supports listed in article.

Institutional review board approval with waiver of written consent (secondary analysis of imaging data from two clinical trials; ClinicalTrials.gov identifiers NCT00756106 and NCT00662506).

Monte Carlo dropout three-dimensional U-Net trained on 64×64×16 voxel patches with dropout probability 0.2 after each convolutional layer; weighted cross-entropy loss; STAPLE used to combine 10 Monte Carlo dropout samples into final segmentation.

Model implemented in the open-source DeepNeuro framework; quantitative metrics computed with the pymia Python package. Data available from corresponding author upon request.

Automated segmentation of postoperative glioblastoma T2-FLAIR abnormality to support expert review and downstream tasks such as treatment planning and response assessment.

Clinical decision support systems; workflow optimization for segmentation review and editing.

Assists decision making by generating automatic segmentations for expert review; not intended for fully autonomous clinical use.

Segmentation of areas of T2-weighted FLAIR abnormality corresponding to total tumor burden in postoperative glioblastoma patients; intended for use by specialists in neuro-oncology, neuroradiology, and radiation oncology within clinical/research environments.

Three-channel MRI: T1-weighted precontrast, T1-weighted postcontrast, and T2-weighted FLAIR sequences (registered to FLAIR) from postoperative glioblastoma patients.

Register T1 pre- and postcontrast to T2-FLAIR; perform brain extraction, N4 bias correction, and z-score normalization of the brain region. Apply the Monte Carlo dropout 3D U-Net with dropout p=0.2 on 64×64×16 patches; aggregate 10 stochastic predictions using STAPLE to obtain the final segmentation.

Study focused on postoperative glioblastoma on T2-FLAIR; findings may not generalize to preoperative tumors or other pathologies. Expert rating used only T2-FLAIR images (no additional sequences during rating). Data derived from two clinical trials; single-manual ground truth per case for training/testing; limited external validation. High interrater variability in perceived segmentation quality; existing metrics (e.g., Dice) correlate poorly with expert perception.

Performance evaluation should include expert-centered assessment and consider metrics aligning with clinical perception (e.g., HD95, surface Dice) rather than relying solely on Dice score.

Processing and evaluation pipelines described (registration, preprocessing, model configuration); implementation using DeepNeuro and pymia. Data available upon request from corresponding author.