Decoding echocardiography: smart detection of view transition using AI and temporal analysis

Abstract Introduction View classification in echocardiographic imaging is essential for accurate analysis and diagnosis. Most automated tools assume that all frames in a sequence belong to the same anatomical view (pure sequence). However, clinical sequences often contain transitions between views (sweep sequence), presenting challenges for reliable measurement and interpretation. While some prior studies briefly acknowledge sweeps [1, 2], none proposes dedicated methods for detecting them. Purpose This study elaborates on current artificial intelligence (AI) view classifiers to explore sweep detection in apical sequences by combining frame prediction with temporal signal analysis. Methods A state-of-the-art AI model [2], trained for frame classification on a clean cohort of 3042 pure apical sequences, was applied to 1042 apical echocardiographic sequences to generate view probability curves (Figure 1). The sequences were annotated 5 times as sweep or pure (ground truth: majority voting) by a cohort of 22 sonographers (mean experience 5±3y). Based on a time-series analysis of the probability curves, a sweep detection algorithm was developed using a forward stepwise grid search to optimize thresholds on the following handcraft engineered features: number of class transitions, ratio of dominant class, sequence length and number of distinct classes. The first 5 frames along with frames with low-confidence predictions (highest probability < 0.6) were discarded. The performance is reported via confusion matrix. The agreement between manual annotations was calculated as reference. Results The model optimization determined two discriminant features, being sweep classified as sequences larger than 15 frames with at least one class transition. The proposed algorithm correctly identified most of the pure sequences (specificity: 83%) and half of the sweep ones (sensitivity: 58%). The inter-annotator agreement was low for sweep sequences, both correctly and incorrectly classified, suggesting that these sequences are intrinsically challenging to interpret and justifying model compromised performance (Figure 2). Misclassified pure sequences had moderate expert agreement, indicating some level of uncertainty, while true pure sequences showed high agreement. The overlapping distributions between sweep and pure sequences across the different features suggest a gradual spectrum, rather than a clear separation, making strict classification inherently difficult (Figure 2). Conclusions This work establishes a proof of concept for automated sweep detection in echocardiography, a poorly addressed challenge in the literature. While performance remains moderate, it is reasonable given the inherent complexity of the task, often overlooked and highlighted in this study by the limited expert agreement. Finally, the proposed approach, combining AI-based frame classification with temporal signal analysis, is directly transferable to alternative AI frame classifiers.Figure 1 Figure 2