Medical Image Computing and Computer Assisted Intervention – MICCAI 2020

Anne L. Martel, Purang Abolmaesumi, Danail Stoyanov, Diana Mateus, María A. Zuluaga, Shuchang Zhou, Daniel Racoceanu, Leo JoskowiczLecture Notes in Computer Science 2020 · DOI · arXiv · OpenAlex

Introduction

Coronary Artery Disease (CAD) remains a primary global health challenge, characterized by arterial narrowing due to plaque accumulation. Clinical assessment relies on the CAD-RADS (Coronary Artery Disease-Reporting and Data System) scoring system, which standardizes reporting from 0 to 5. However, manual calculation is labor-intensive and susceptible to inter-observer variability. This research presents a deep-learning-based "bottom-up" approach to automate CAD-RADS scoring from coronary CT scans.

Methodology

The proposed system utilizes a hierarchical, multi-task deep learning architecture designed to mimic clinical workflows.

Architecture and Representation

The process begins with automated centerline extraction and anatomical labeling of vessel segments. To balance computational efficiency with spatial detail, the model employs 2.5D Multi-Planar Reformatted (MPR) image stacks. A shared-weight CNN extracts features from each segment, which are then aggregated via a global max pooling layer to identify the most severe stenosis across the entire coronary tree.

Hierarchical multi-task architecture for CAD-RADS scoring.

Multi-Task Learning and Regression

The network is trained to simultaneously predict segment-wise stenosis and patient-level calcification grades. By treating CAD-RADS scoring as a regression task rather than simple classification, the model respects the ordinal nature of the disease progression, penalizing larger deviations more heavily.

Experimental Setup

The study leveraged a large dataset of 2,867 patients, with 968 reserved for testing. Labels included segment-level stenosis, patient-level CAD-RADS scores, and Agatston scores for calcium burden. Patients with existing stents or bypass grafts were excluded to focus on primary diagnosis.

Results and Discussion

The multi-task approach significantly outperformed baseline models. The inclusion of auxiliary tasks like calcification prediction provided a strong "rule-out" signal, enhancing overall diagnostic accuracy.

ROC curves for Rule-Out, Hold-Out, and Significant Stenosis tasks.

Key performance metrics include:

  • Rule-Out (CAD-RADS 0 vs. 1-5): AUC of 0.914.
  • Hold-Out (CAD-RADS 0-2 vs. 3-5): AUC of 0.923.
  • Overall Accuracy: Improved from 0.810 to 0.840 with the full multi-task configuration.

Conclusion

This automated "bottom-up" framework demonstrates high diagnostic performance suitable for clinical screening. By providing segment-level insights alongside a global patient score, the model offers interpretability that supports radiologist decision-making. Future work will focus on incorporating high-risk plaque modifiers and extending the model to handle post-intervention cases like stents and bypasses.

Summary updated

BibTeX
@inproceedings{martel2020medical,
  author = {Anne L. Martel and Purang Abolmaesumi and Danail Stoyanov and Diana Mateus and María A. Zuluaga and Shuchang Zhou and Daniel Racoceanu and Leo Joskowicz},
  title = {{Medical Image Computing and Computer Assisted Intervention – MICCAI 2020}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2020},
  doi = {10.1007/978-3-030-59725-2},
}
Figure