Improved image quality and greater diagnostic suitability in myocardial delayed enhancement CT with deep learning image reconstruction

Myocardial delayed enhancement computed tomography (MDE-CT) is an emerging imaging modality for assessing myocardial fibrosis. However, its diagnostic performance is often limited by low contrast resolution and high image noise. Deep learning–based image reconstruction (DLIR) has recently been introduced as a novel method to enhance CT image quality. This study aimed to evaluate whether DLIR improves image quality and diagnostic suitability in MDE-CT, compared to conventional hybrid iterative reconstruction (HIR). A total of 108 patients with visually confirmed myocardial delayed enhancement on CT were included. CT images were reconstructed using both HIR and DLIR. Quantitative image quality metrics included image noise, contrast-to-noise ratio (CNR), and signal-to-noise ratio (SNR). Qualitative image quality was independently assessed by two radiologists using a 5-point Likert scale (1 = poor, 5 = excellent), with scores ≥ 3 considered diagnostically suitable. DLIR significantly reduced image noise (median 7.1 Hounsfield unit [HU] vs. 9.2 HU) and improved both CNR (median 3.2 vs. 2.6) and SNR (median 11.7 vs. 9.0) compared to HIR (all p < 0.0001). DLIR increased CNR and SNR by 26.9% and 27.1%, respectively. Qualitative scores were also significantly higher for DLIR (Observer 1: 4.2 ± 0.8 vs. 3.4 ± 0.8; Observer 2: 3.6 ± 0.8 vs. 3.2 ± 0.9; all p < 0.0001). The proportion of diagnostically suitable images significantly increased in both readers (Observer 1: 88.9% [96/108] to 97.2% [105/108]; Observer 2: 82.4% [89/108] to 90.7% [98/108]; both p < 0.03). DLIR significantly improves both quantitative and qualitative image quality in MDE-CT, resulting in a higher proportion of diagnostically suitable images. These improvements support the incorporation of DLIR into routine MDE-CT protocols as a robust alternative to conventional iterative reconstruction.