Deep-learning tool for early identification of non-traumatic intracranial hemorrhage etiology and application in clinical diagnostics based on computed tomography (CT) scans

Meng Zhao, Wenjie Li, Yifan Hu, Ruixuan Jiang, Yuanli Zhao, Dong Zhang, Yan Zhang, Rong Wang, Yong Cao, Qian Zhang, Yonggang Ma, Jiaxi Li, Shaochen Yu, Ran Zhang, Yefeng Zheng, Shuo Wang, Jizong ZhaoPeerJ 2025 · DOI

TL;DR

This paper introduces ICHNet, a deep learning system that accurately identifies the cause of non-traumatic intracranial hemorrhage (ICH) from standard CT scans. The system not only matches expert clinician performance but also significantly improves their diagnostic accuracy when used as a supportive tool, paving the way for faster and more precise patient triage and treatment.

What Problem Does This Paper Address?

Intracranial hemorrhage (ICH) is a major cause of stroke-related death and disability. While hypertension is the most common cause, a significant portion of ICH cases stem from underlying vascular abnormalities like aneurysms, arteriovenous malformations (AVMs), and Moyamoya disease (MMD). Accurately identifying the specific cause of ICH is critical because treatment strategies vary widely. Current diagnostic pathways often rely on invasive or expensive imaging techniques like CT angiography (CTA) or magnetic resonance angiography (MRA), which may not be readily available in all clinical settings, especially in the acute phase. Non-contrast computed tomography (NCCT) is the initial imaging modality, but extracting detailed etiological information from it is challenging. This paper aims to bridge this gap by developing a deep learning model that can accurately predict ICH etiology directly from NCCT scans, thereby aiding clinicians in making faster, more informed decisions.

Methods

The researchers developed a deep learning model named ICHNet, inspired by the CNN architecture SlowFast Networks. The model was trained on a large dataset of NCCT scans from patients with non-traumatic ICH. Key preprocessing steps included rotating the volumes, resizing them to a consistent voxel size, and applying an unsupervised intensity-based skull-stripping algorithm. The model was trained using a combination of weighted categorical cross-entropy loss and triplet loss to handle class imbalance and learn robust representations.

For evaluation, the model was tested on two independent datasets. Crucially, the study also simulated a clinical setting by comparing the performance of six expert raters (neuroradiologists and neurosurgeons) in three tasks:

  1. NCCT images only: Raters diagnosed solely based on CT scans.
  2. NCCT + clinical data: Raters had access to patient demographics, medical history, etc.
  3. NCCT + clinical data + model predictions: Raters could use the deep learning model's probabilistic predictions.

This multi-task evaluation allowed the researchers to assess the model's standalone performance and its impact on clinician decision-making.

Datasets

  • Development Dataset: 1,868 eligible NCCT scans from patients at Beijing Tiantan Hospital (January 2011 - April 2018). This dataset was used to train and validate the ICHNet model.
  • TT200 Dataset: 200 NCCT scans from patients at Beijing Tiantan Hospital (April 2018 - November 2018). This served as an independent test set.
  • SD98 Dataset: 98 NCCT scans from patients at the Affiliated Hospital of Shandong Jining Medical College (April 2018 - January 2019). This served as a second independent test set, adding diversity.

The paper also mentions that the raw data measurements are available in a supplemental file and on Zenodo, and the code is available on GitHub.

Key Insights and Ideas

  • NCCT is Richer Than We Thought: The study demonstrates that standard NCCT scans, often considered primarily for hemorrhage detection, contain enough information for a deep learning model to accurately infer the underlying etiology of ICH.
  • AI as a Clinician Augmenter, Not Replacer: The most significant finding is that while the deep learning model performs comparably to experts, its true value lies in its ability to enhance clinician performance. When clinicians had access to the model's predictions, their diagnostic accuracy, sensitivity, and specificity improved significantly.
  • Addressing Class Imbalance is Crucial: The researchers employed oversampling techniques for minority classes (like MMD and CM) during training, highlighting the importance of this strategy for building robust models for rare conditions.
  • Concordance Matters: The study showed a substantial increase in inter-rater agreement (measured by Cohen's kappa) after clinicians used the AI system, suggesting it leads to more consistent diagnoses.

Key Results

  • Model Performance: ICHNet achieved high Area Under the Receiver Operating Characteristic Curve (AUC) values across various etiologies on the TT200 test dataset, including 0.986 for aneurysms and 0.952 for hypertensive hemorrhage. Similar strong performance was observed on the SD98 dataset.
  • Clinician Performance Improvement:
    • Task 1 (NCCT only): The model's accuracy (0.760) outperformed the average clinician accuracy (0.706).
  • Task 2 (NCCT + clinical data): Clinician accuracy improved to 0.725, but the model still performed better.
  • Task 3 (NCCT + clinical data + AI): Clinicians showed significant improvements in sensitivity and specificity for several etiologies, with overall accuracy increasing from 0.725 to 0.803 (p < 0.01).
  • Inter-Rater Agreement: Cohen's kappa increased from a moderate range (0.51-0.69) before AI assistance to an excellent range (0.83, p < 0.01) after, with 9 out of 15 pairs showing excellent agreement.

Limitations and Future Work

  • Dataset Bias: The study was hospital-based, and the datasets were from large tertiary referral hospitals. This may introduce bias towards more complex or referral cases, and the distribution of etiologies was unbalanced, with limited samples for minority classes like MMD and CM.
  • Generalizability: While two independent datasets were used, further validation on more diverse, multi-center datasets is needed.
  • Clinical Integration: The authors emphasize the need for careful consideration and validation before integrating this tool into clinical workflows, focusing on patient safety and avoiding issues like automation bias.
  • Future Directions:
    • Enriching the training database, especially for minority classes.
  • Exploring the use of spectral CT data for potentially improved diagnostic accuracy.
  • Developing explainable AI (XAI) techniques to increase clinician trust and facilitate error detection.
  • Investigating how to integrate this tool into clinical workflows to optimize patient care outcomes.

Summary updated

Abstract

The proposed deep learning tool can be an accuracy tool for early identification of hemorrhage etiologies based on NCCT scans. It may also provide more information for clinicians for triage and further imaging examination selection.
BibTeX
@inproceedings{zhao2025deep,
  author = {Meng Zhao and Wenjie Li and Yifan Hu and Ruixuan Jiang and Yuanli Zhao and Dong Zhang and Yan Zhang and Rong Wang and Yong Cao and Qian Zhang and Yonggang Ma and Jiaxi Li and Shaochen Yu and Ran Zhang and Yefeng Zheng and Shuo Wang and Jizong Zhao},
  title = {{Deep-learning tool for early identification of non-traumatic intracranial hemorrhage etiology and application in clinical diagnostics based on computed tomography (CT) scans}},
  booktitle = {PeerJ},
  year = {2025},
  doi = {10.7717/peerj.18850},
}
Figure