Hybrid Model for Brain Hemorrhage Segmentation
Author Information
Author(s): Tran Anh T., Desser Dmitriy, Zeevi Tal, Abou Karam Gaby, Dierksen Fiona, Dell’Orco Andrea, Kniep Helge, Hanning Uta, Fiehler Jens, Zietz Julia, Sanelli Pina C., Malhotra Ajay, Duncan James S., Aneja Sanjay, Falcone Guido J., Qureshi Adnan I., Sheth Kevin N., Nawabi Jawed, Payabvash Seyedmehdi, Aricò Pietro
Primary Institution: Yale School of Medicine
Hypothesis
Can a hybrid transformer-convolutional neural network improve segmentation of intracerebral hemorrhage and perihematomal edema on CT scans?
Conclusion
The dual Swin-nnUNET model significantly improved segmentation accuracy for intracerebral hemorrhage and perihematomal edema compared to traditional methods.
Supporting Evidence
- The model achieved a median Dice coefficient of 0.93 for ICH segmentation in the Yale cohort.
- Segmentation accuracy was significantly higher in the dual Swin-nnUNET model compared to traditional methods.
- Prediction uncertainty was associated with lower segmentation accuracy and smaller hemorrhage volumes.
- The model was validated on independent datasets from Yale and Charité University Hospital.
- Uncertainty detection improved overall segmentation performance by filtering out low-confidence predictions.
Takeaway
This study created a smart computer program that helps doctors quickly and accurately find brain bleeding and swelling in CT scans, which can help treat patients better.
Methodology
The study used a hybrid model combining Swin-Unet Transformers and nnU-NETv2 convolutional networks, trained on 1782 CT scans, and evaluated using Dice coefficient and Volume Similarity metrics.
Potential Biases
Manual segmentation differences between institutions could introduce variability in model training and testing.
Limitations
The model's performance may be suboptimal for scans obtained beyond 36 hours after ICH onset, and it may struggle with infratentorial hemorrhages due to under-representation in training data.
Participant Demographics
CT scans were from patients with acute spontaneous intracerebral hemorrhage, with specific inclusion criteria based on time from onset and hematoma volume.
Statistical Information
P-Value
p<0.001
Confidence Interval
95% CI for Dice = 0.90–0.95
Statistical Significance
p<0.001
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website