TL;DR
This paper introduces HemSeg-200, a novel, voxel-level annotated dataset of 222 brain CT scans for segmenting intracerebral hemorrhages (IPH and IVH). This dataset aims to bridge the gap in quantitative analysis for these life-threatening conditions and was used to evaluate seven state-of-the-art segmentation models, with nnU-Net achieving the best performance.
What Problem Does This Paper Address?
Acute intracerebral hemorrhage (ICH) is a severe medical emergency with high mortality rates. Accurate assessment of bleeding extent, particularly intraparenchymal hemorrhage (IPH) and intraventricular hemorrhage (IVH), is crucial for treatment planning. While deep learning has shown promise in diagnosing ICH, there's a significant need for quantitative analysis, which requires precise segmentation of these hemorrhages. Existing public datasets often lack voxel-level annotations or are limited in size, hindering the development and evaluation of advanced segmentation models for this critical task.
Methods
The authors developed a semi-automated, coarse-to-fine annotation strategy to create their dataset. This involved using a pre-trained tool (TotalSegmentator) for initial segmentation, followed by independent annotation by junior annotators, and finally, a thorough review and correction by a board-certified neuroradiologist. For model evaluation, they tested seven commonly used 3D medical image segmentation architectures: 3D U-Net, V-Net, SegResNet, Attention U-Net, UNETR, Swin UNETR, and nnU-Net. The Dice loss function was used for training.
Datasets
The primary contribution is the HemSeg-200 dataset, comprising 222 brain CT volumes. Of these, 114 are annotated for IPH and 108 for IVH. The data was sourced from the RSNA 2019 Brain CT Hemorrhage Challenge dataset. The authors highlight that HemSeg-200 is unique in providing voxel-level annotations, addressing a key limitation of existing datasets like RSNA and CQ500, which lack such detailed annotations. They also note that other public datasets like PhysioNet and INSTANCE22 are limited in sample size.
Key Insights and Ideas
- The critical need for voxel-level annotations: The paper strongly emphasizes that for quantitative analysis and robust model development in medical image segmentation, precise voxel-level annotations are indispensable, a gap that HemSeg-200 aims to fill.
- Semi-automated annotation is feasible and effective: The proposed multi-step annotation process, combining automated tools with expert review, demonstrates a practical approach to creating large, high-quality annotated medical datasets, which are notoriously labor-intensive.
- nnU-Net's robustness: The consistent high performance of nnU-net across various metrics suggests its self-configuring nature is highly effective for medical image segmentation tasks, even on novel datasets.
- Transformer limitations on this specific data: Despite the success of transformer-based models in other domains, UNETR and Swin UNETR did not perform as well on this dataset, indicating that dataset characteristics (e.g., hemorrhage heterogeneity) can significantly impact model performance.
Key Results
The HemSeg-200 dataset was used to evaluate seven 3D segmentation models. nnU-Net achieved the highest performance, with a Dice score of 78.99% and an IoU of 68.49% on the independent test set. The confidence intervals for nnU-Net's Dice score did not overlap with those of other models, indicating a statistically significant advantage. Attention U-Net performed second best, while V-Net showed the best performance in terms of HD95 (48.88). Transformer-based models like UNETR and Swin UNETR showed lower performance compared to U-Net-based architectures.
Limitations and Future Work
The authors acknowledge that despite nnU-Net's strong performance, the overall segmentation accuracy is still limited, suggesting the need for more data. Their future work includes continuing the annotation process to develop even larger datasets. They also plan to compare nnU-Net with other autoconfigured models and explore more advanced transformer-based architectures. The paper also notes that the dataset may contain a mixture of new and old hemorrhages, as well as ICH and IVH, which can pose challenges for segmentation algorithms.
Summary updated