PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 26, 2025Frontiers in Medicine2 citationsOpen Access

Beyond plaque segmentation: a combined radiomics-deep learning approach for automated CAD-RADS classification

View Full Paper
Ask AI
Bookmark
Share

Why the study?

Accurate CAD stenosis grading requires time-consuming manual assessment and suffers from interobserver variability, and no deep learning methods had explored combining radiomic and autoencoder-based features.

Does a combined radiomics-deep learning approach improve automated CAD-RADS classification accuracy in patients with suspected CAD compared to single-domain models?

Population

2,548 CCTA-derived MPR images from 220 patients

Comparison

Combined radiomic and AE-based features vs single feature sets

Design

Machine learning model development and validation study

Key result

A combined radiomics and autoencoder-based machine learning model achieved an overall balanced accuracy of 0.91 for automated CAD-RADS classification, outperforming single-domain models.

Authors

FIFrancesca Lo IaconoFRFrancesca RonchettiACAnna Corti

Discussion

Loading...

Member takes

Overview

May support automated CAD-RADS scoring in suspected CAD; hypothesis-generating and requires prospective validation before clinical adoption.

Study Design

Type

Observational (n=220)

Multicenter

No

Structured PICO

Does a combined radiomics-deep learning approach improve automated CAD-RADS classification accuracy in patients with suspected CAD compared to single-domain models?

P
Population
220 patients with suspected coronary artery disease (CAD) who underwent CCTA, classified as no-CAD (n=40), non-obstructive CAD (n=80), or obstructive CAD (n=100). Mean age 60.8 years, 70% male, based in Italy.
I
Intervention
Combined machine learning approach using 64 autoencoder (AE)-based and 465 2D radiomic features extracted from CCTA-derived multiplanar reconstructed (MPR) images.
C
Comparator
Single models using either AE-based features alone or radiomic features alone.
O
Outcome
Overall balanced accuracy, sensitivity, and specificity for stratifying patients into no-CAD, non-obstructive CAD, or obstructive CAD on the test set.

Combining radiomic and autoencoder-based deep learning features from CCTA images enables highly accurate, automated patient-level CAD-RADS classification, potentially reducing manual assessment time and interobserver variability.

Main Result

Absolute Event Rate: 0.91% vs 0.82%

Limitations

  • Significant class imbalance in the dataset
  • Single-center nature of the data collection
  • Absence of external validation
  • Challenges in exhaustive interpretation of the AE-derived representation and direct clinical correlations

Cite This Study

Iacono et al. (2025) conducted an observational in Coronary Artery Disease (n=220). Combined radiomics and autoencoder (AE)-based machine learning model vs. Single radiomic model and single AE-based model was evaluated on Overall balanced accuracy for 3-class CAD-RADS stratification on the test set. A combined radiomics and autoencoder-based machine learning model achieved an overall balanced accuracy of 0.91 for automated CAD-RADS classification, outperforming single-domain models.

synapsesocial.com/papers/6a0ee4cb53f874f2b222e9cchttps://doi.org/10.3389/fmed.2025.1536239
View Full Paper
Ask AI
Bookmark
Share