Semantic Segmentation and Pathology Localization in Lung Ultrasound Images Using Transfer Learning

Significant progress has been made to leverage
machine learning towards decision support diagnosis from
ultrasound in medical imaging. Due to the continuous challenges
associated with evolving characteristics of COVID-19, the
availability of fast, safe, and highly sensitive diagnostic tools is
imperative. Where ultrasound imaging has been proven superior
to X-rays and CT scans, the main limitation to its use is operator
dependency and experience. In this study, we propose a novel
transfer learning and semantic segmentation framework for
automatic detection and localization of multiple lung pathologies
in Lung Ultrasound images. The proposed framework allows
better interpretation and explanation of the model decisions, with
clear visualization and localization of the different pathologies.
The experimental evaluation of the proposed framework was done
on an open-source Lung ultrasound imaging dataset which is
labeled and annotated by team of radiologists. The proposed
approach was validated using other benchmarks models for
comparison in terms of DICE coefficients at 0.98 and IOU Score
at 0.97 which outperforms the other benchmark models quite
significantly. This provides interpretation of the reasoning behind
the decision made by the model leading to higher rates of
acceptance with clinicians.