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

Focal loss 3D nnU-Net for ICH/PHE/IVH CT segmentation (TICH-2)

https://dx.doi.org/10.1148/ryai.220096

1.0

This study included data from the TICH-2 trial funded in part by the UK National Institute for Health Research Health Technology Assessment Programme (project code 11_129_109).

Ethical approval was granted from the UK Health Research Authority and relevant national or local institutional review boards; written informed consent was obtained before enrollment.

Three-dimensional nnU-Net (U-Net–based convolutional neural network); comparisons included 2D nnU-Net, ensemble (2D+3D), U-Net (MONAI BasicUNet), BLAST-CT (DeepMedic-based), and DeepLabv3+.

Implementations used: nnU-Net (https://github.com/MIC-DKFZ/nnUNet); BLAST-CT (https://github.com/biomedia-mira/blast-ct); MONAI BasicUNet; DeepLabv3+ implementation used by authors (https://github.com/janetkok/MONAI-DeepLabV3plus); loss function code (https://github.com/JunMa11/SegLoss.git). Data sharing on request to the TICH-2 Chief Investigator as per article.

Automated segmentation and quantification of ICH, PHE, and IVH to provide quantitative outcome measures for clinical trials and potentially accelerate studies in large cohorts of spontaneous ICH.

Clinical research/clinical trials biomarker quantification; potential decision support for outcome prediction studies.

Fully automated image segmentation of target lesions on noncontrast head CT.

Segmentation and volumetric quantification of intracerebral hemorrhage, perihematomal edema, and intraventricular hemorrhage on baseline noncontrast CT in adults with spontaneous ICH (within 8 hours of symptom onset in the TICH-2 cohort).

Baseline noncontrast head CT scans from patients with spontaneous intracerebral hemorrhage.

Models trained for 1800 epochs using default pipeline parameters of respective frameworks; three-dimensional nnU-Net refined with Focal/Tversky/FocalTversky/DiceTopK loss functions to address class imbalance.

Lower performance and smoothed boundaries for PHE segmentation; single train/test split (no k-fold cross-validation); DSC can be inflated by cases with absent lesions; ground truth contains some segmentation errors; class imbalance with approximately 30% IVH prevalence; multicenter CT variability.

U-Net–based networks, particularly 3D nnU-Net with Focal loss, are recommended to address class imbalance and improve IVH segmentation; performance satisfactory for ICH and moderate for PHE.

Open-source frameworks were used (nnU-Net, MONAI, BLAST-CT); models trained with 1800 epochs; dataset split 90/10; specifics detailed in the supplement (Appendix E1/Table E1).