Active learning for left ventricle segmentation in echocardiography

• DOI: 10.1016/j.cmpb.2024.108111
• Publication Date: 2024-03-07T17:08:03Z
• Authors: Eman Alajrami; Tiffany Ng; Jevgeni Jevsikov; Preshen Naidoo; Patricia Fernandes; Neda Azarmehr; Fateme Dinmohammadi; Matthew J. Shun-shin; Nasim Dadashi Serej; Darrel P. Francis; Massoud Zolgharni
• Keywords: Feature (linguistics); Baseline (sea)

Background and Objective: Training deep learning models for medical image segmentation requires large annotated datasets, which can be expensive time-consuming to create. Active is a promising approach reduce this burden by strategically selecting the most informative samples segmentation. This study investigates use of active efficient left ventricle in echocardiography with sparse expert annotations. Methods: We adapt evaluate various sampling techniques, demonstrating their effectiveness judiciously Additionally, we introduce novel strategy, Optimised Representativeness Sampling, combines feature-based outliers representative enhance annotation efficiency. Results: Our findings demonstrate substantial reduction costs, achieving remarkable 99% upper bound performance while utilizing only 20% labelled data. equates 1680 images needing within our dataset. When applied publicly available dataset, yielded 70% required efforts, representing significant advancement compared baseline strategies, achieved 50% reduction. experiments highlight nuanced diverse strategies across datasets same domain. Conclusions: The provides cost-effective tackle challenges limited annotations echocardiography. By introducing distinct made research purposes, work contributes field's understanding