An interpretable artificial intelligence model based on CT for prognosis of intracerebral hemorrhage: a multicenter study

Hao Zhang, Yun-Feng Yang, Xue-Lin Song, Hai-Jian Hu, Yuan-Yuan Yang, Xia Zhu, Chao YangBMC Medical Imaging 2024 · DOI

TL;DR

This paper introduces a novel, interpretable AI model that combines radiomic, deep learning, and clinical features to predict the prognosis of intracerebral hemorrhage (ICH) patients. By integrating these diverse data sources and employing explainability techniques, the model achieves high predictive accuracy while offering insights into its decision-making process, paving the way for more trustworthy AI in clinical settings.

What Problem Does This Paper Address?

Intracerebral hemorrhage (ICH) is a severe condition with high mortality and disability rates. Accurately predicting patient prognosis is crucial for timely and effective treatment, but current methods often lack precision or interpretability. While AI has shown promise in medical imaging, many high-performing models are "black boxes," making it difficult for clinicians to trust their predictions. This study aims to bridge this gap by developing an AI model that is both accurate and interpretable, integrating various data modalities to improve prognostic prediction for ICH patients.

Methods

The researchers developed a multimodal AI model by combining three types of features:

  • Radiomic Features: Extracted using the Pyradiomics library, these features quantify various characteristics of the hematoma from CT images.
  • Deep Learning Features: A pre-trained EfficientNetV2 model was fine-tuned to extract 1000 deep learning features from the CT images, capturing complex patterns.
  • Clinical Features: Standard clinical data such as age, sex, and comorbidities were included.
  • Imaging Features: Radiologists identified and evaluated specific qualitative imaging features from the CT scans.

These features were then fused and fed into a Random Forest (RF) classifier. To ensure interpretability, the study employed:

  • SHAP (Shapley Additive Explanations): A game-theory-based method to quantify the contribution of each feature to the model's predictions.
  • Grad-CAM and Guided Grad-CAM: Visualization techniques to highlight the regions in the CT images that the deep learning model focused on for its predictions.

The model was trained and validated on data from two medical centers to ensure robustness.

Datasets

The study utilized a retrospective dataset of 222 patients with ICH from two medical centers:

  • Training Cohort: 186 patients from Medical Center 1.
  • External Testing Cohort: 36 patients from Medical Center 2.

The data included Non-contrast Computed Tomography (NCCT) images and clinical data. The outcome was assessed at six months post-onset using the modified Rankin Scale (mRS), with scores 0-2 considered a "good outcome" and 3-6 a "poor outcome."

Key Insights and Ideas

  • Multimodal Fusion is Key: Combining radiomic, deep learning, and clinical features significantly outperforms models relying on single modalities, highlighting the complementary nature of different data types.
  • Interpretability Enhances Trust: The integration of SHAP, Grad-CAM, and Guided Grad-CAM provides crucial insights into why the model makes certain predictions, making it more acceptable for clinical use.
  • Novel Composite Feature: The introduction of a "DL-Clinical-Rad-Score" as a fused feature demonstrated a substantial contribution to the model's overall decision-making.
  • Radiomics Captures Nuance: Specific radiomic features related to image grayscale levels and shape, as well as texture features, were found to be highly influential in predicting prognosis.

Key Results

The developed DL-Radiomics-Clinical Model achieved the highest predictive performance on the external testing cohort:

  • Area Under the Receiver Operating Characteristic Curve (AUC): 0.89 (95% CI: 0.74, 0.97; P<0.01).
  • The Radiomics-Clinical Model also showed strong performance with an AUC of 0.86 (95% CI: 0.71, 0.95; P<0.01).
  • The Radiomics Model alone achieved an AUC of 0.83 (95% CI: 0.67, 0.93; P<0.01).
  • The study found no statistically significant differences in ROC curves between the models when using the Delong test, but the DL-Radiomics-Clinical Model showed the best generalization.
  • SHAP analysis revealed that the fused DL-Clinical-Rad-Score was the most significant contributor to the model's predictions.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Data Expansion: While multicenter data was used, including more centers and larger datasets could further improve generalization and accuracy, especially for complex AI algorithms.
  • Fusion Method: The study used an "early fusion" approach. Exploring more advanced multimodal feature fusion techniques could potentially enhance model effectiveness.
  • Hematoma Expansion: The study acknowledges that hematoma expansion is a critical factor for prognosis but was limited by the availability of repeat CT scans in the dataset. Future studies should incorporate clinical evidence of hematoma expansion.
  • Interpretability Depth: While SHAP provides quantitative insights, visual interpretability methods like Grad-CAM and Guided Grad-CAM, though useful, cannot fully quantify specific feature impacts. Future research could explore methods offering higher accuracy and interpretability.
  • Model Deployment: The developed RF model is lightweight, requiring minimal hardware, which is beneficial for easy deployment. However, further research into optimizing deep learning models for efficiency and deployment on resource-constrained devices is ongoing.

Summary updated

BibTeX
@article{zhang2024interpretable,
  author = {Hao Zhang and Yun-Feng Yang and Xue-Lin Song and Hai-Jian Hu and Yuan-Yuan Yang and Xia Zhu and Chao Yang},
  title = {{An interpretable artificial intelligence model based on CT for prognosis of intracerebral hemorrhage: a multicenter study}},
  journal = {BMC Medical Imaging},
  year = {2024},
  doi = {10.1186/s12880-024-01352-y},
}
Figure