EchoVisuAL: Efficient Segmentation of Echocardiograms Using Deep Active Learning

Echocardiography is a fast and cost-effective imaging technique for assessing cardiac function and structure. However, image-derived phenotypic evaluation is challenging. Current AI-approaches designed for automatic interpretation of echocardiography data are progressing, but algorithms for animal models frequently used in pre-clinical studies are rare. Here, we propose a deep active learning approach, called EchoVisuAL, that uses large-scale, multi-center data of the International Mouse Phenotyping Consortium (IMPC). This heterogeneous IMPC data set includes 96 392 echocardiograms with 3 831 290 frames from 17 991 mice. Heterogeneity is characterized by differences in age, sex, background strains, anesthesia, imaging frequency and focus depth. EchoVisuAL is founded on a Bayesian U-Net that produces inner trace segmentations alongside with two confidence metrics, an uncertainty measure and a BALD score. This architecture, embedded in an active learning framework, enables a substantial reduction of the annotation efforts by an intelligent selection of the next frames that should be annotated. In total, 15 models were trained on step-wise increasing training data sets based on the model’s confidence. For model evaluation, 25 echocardiograms with 1062 frames were annotated by four highly experienced, independent experts. Inter-rater-agreement across all frames was high with a mean Randolph’s kappa score of \(0.91\pm 0.10\) . Across models, high Dice scores were observed on these expert annotations, currently considered as the gold standard, with model \(M_{15}\) achieving a mean Dice score of \(0.98\pm 0.02\) . EchoVisuAL is a new deep active learning application robust to automatically analyze heterogeneous mouse echocardiograms, including uncertainty scores for user guidance.