Deep-Learning Model for Quality Assessment of Urinary Bladder Ultrasound Images Using Multiscale and Higher-Order...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous ultrasound image quality assessment (US-IQA) is crucial for clinical interpretation and robotic procedures.
• Challenges in US-IQA include image artifacts (noise, probe positioning errors) and patient-specific anatomical variations.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN), USQNet, for sonographer-like autonomous ultrasound image quality assessment.
• To address challenges in US-IQA by utilizing a multiscale and local-to-global second-order pooling (MS-L2GSoP) classifier.

Main Methods:

• USQNet employs a CNN architecture with a novel MS-L2GSoP classifier.
• The MS-L2GSoP classifier extracts multi-scale features for anatomical variations and uses second-order pooling (SoP) to capture statistical dependencies.
• Validation was performed on a new dataset of human urinary bladder ultrasound images, comparing against radiologist assessments and state-of-the-art CNNs.

Main Results:

• USQNet achieved a high accuracy of 92.4% in autonomous ultrasound image quality assessment.
• The model outperformed existing state-of-the-art CNNs for US-IQA by 3%-14%.
• USQNet demonstrated comparable computational time to other models.

Conclusions:

• USQNet offers a robust and accurate solution for autonomous ultrasound image quality assessment.
• The developed model can assist sonographers and advance the robotization of ultrasound procedures.
• The MS-L2GSoP classifier effectively captures image features for reliable US-IQA.