Radiomics and deep learning for myocardial scar screening in hypertrophic cardiomyopathy

Area of Science:

• Cardiovascular Magnetic Resonance (CMR) Imaging
• Artificial Intelligence (AI) in Medical Diagnostics
• Cardiology and Cardiovascular Diseases

Background:

• Late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) quantifies myocardial scar burden, crucial for hypertrophic cardiomyopathy (HCM) prognosis.
• Approximately 50% of HCM patients lack scar, yet undergo repeated gadolinium-based CMR scans throughout their lives.
• There is a need for non-contrast methods to identify HCM patients without scar, avoiding repeated gadolinium administration.

Purpose of the Study:

• To develop an artificial intelligence (AI)-based screening model to identify HCM patients without myocardial scar.
• To utilize radiomics and deep learning (DL) features from balanced steady state free precession (bSSFP) cine sequences for scar screening.
• To reduce the need for gadolinium-based contrast agents in routine HCM follow-up.

Main Methods:

• Three AI models were evaluated: radiomics, deep learning (DL), and a combined DL-Radiomics approach.
• Models were trained and tested on 759 HCM patients from a multi-center/vendor study using bSSFP cine images.
• Model generalizability was assessed using an external dataset of 100 HCM patients; performance was measured by area-under-receiver-operating curve (AUC).

Main Results:

• The DL-Radiomics model achieved higher AUC than DL or radiomics alone in both internal (0.83 vs. 0.77/0.78) and external (0.74 vs. 0.64/0.71) datasets.
• The DL-Radiomics model correctly identified 43% (internal) and 28% (external) of patients without scar.
• Radiomics and DL models identified 42% (internal) and 16-23% (external) of scar-negative patients, respectively.

Conclusions:

• An AI model integrating DL and radiomics features from bSSFP cine images serves as an effective scar screening tool prior to gadolinium administration in HCM.
• This combined AI approach surpasses individual DL or radiomics models for scar screening in HCM patients.
• Further research is necessary to enhance the accuracy and generalizability of the AI model for broader clinical utility.