Enhancing image quality quantification in echocardiography by combining multiple quality dimensions

Abstract Introduction Echocardiographic image quality is crucial for accurate cardiac assessment [1], yet its evaluation is inherently subjective [2]. Furthermore, the lack of a standard criterion adds to this challenge, as current methods vary widely in both their levels of granularity and the scope of questions addressed [3,4]. Purpose This study aims to refine echocardiographic image quality assessment by laying the groundwork for an objective and informative scoring system that enhances diagnostic reliability. To achieve this, we deconstruct image quality into five distinct dimensions, exploring both their individual contributions and their potential synergy when combined into a composite metric. Methods A total of 15,712 transthoracic echocardiograms from 888 patients were evaluated for image quality by a panel of 23 experienced sonographers (mean expertise 5 ± 3 years). For each sequence, a 5-point Likert scale (1=lowest, 5=highest) was applied across five domains: sharpness, artifact absence, acquisition plane, contour visibility, and validity for clinical measurements. To explore the underlying structure in these ordinal ratings, we computed polychoric correlations and performed an eigenvalue analysis to determine latent dimensionality. Graded item response theory (IRT) was applied to assess how each domain relates to these latent traits. In essence, this psychometric method—widely used in patient-reported outcomes yet novel for image quality—assesses each domain’s discrimination and yields a composite score. Results Our analysis indicates that the five-domain scale measures indeed a single dominant latent trait, likely representing overall image quality, as confirmed by eigenvalue analysis (Figure 1A). However, item probability functions (Figure 2A) revealed that the five quality dimensions often received similar ratings, offering little unique information to capture this latent trait. Peaks in the IRT test information function at Likert transitions (Figure 1B) imply that responses frequently cluster, limiting scale granularity. Individual item information (Figure 2B) shows significant differences in item utility: "contour visibility" was the most informative and "clinical validity" showed moderate discrimination—likely due to its dependence on upstream quality factors. The other three domains contributed minimally, suggesting their definitions may be ambiguous or their sensitivity low to discriminate between levels of image quality. Conclusion Thanks to the herein proposed novel framework to assess image quality, we discovered that the five quality dimensions we assessed essentially capture a single underlying aspect, which supports the idea of combining them into one overall score. However, we also found considerable overlap among these measures. Future work should leverage this framework to refine quality definitions, confirm clinical relevance, and ultimately set a standardized unique quality criterion for the community.Figure 1 Figure 2