Development of A Real-time POCUS Image Quality Assessment and Acquisition Guidance System

Automated acquisition guidance can help novice learners to study how to manipulate probes to acquire high-quality POCUS images. However, the difference between human vision and machine vision may confuse novice learners, because the guidance from sonographers and the DNN model may be inconsistent due to the different interpretations of images between the DNN-based models and humans [6]. For example, DNNs rely more on textures and high-frequency features rather than shapes and low-frequency features. Our previous works show that machine learning based detection methods on medical image can achieve significantly high accuracy [7-15]. This means images considered highquality for humans may not be high-quality for DNN-based machine vision. The discrepancy becomes more significant due to the large inter-observer variability. As a result, if the acquired images are of high quality for both human and DNN-based segmentation models, the segmentation accuracy can be improved, and accordingly, the human efforts needed to correct or identify the bad segmentation can be reduced. As an example in a closely-related area, previous works for medical image compression demonstrated that image compressed with machine vision considered achieved significantly higher segmentation accuracy than conventional methods that only consider human vision at the same compression rate [16]. Therefore, we will explore methods to optimize the image quality assessment and acquisition guidance to be preferred by both human and machine models.