ClinReadNet: A clinical reading-inspired network for low-dose abdominal CT image quality assessment

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate abdominal CT image quality assessment is vital for balancing radiation dose and image quality.
• Current subjective assessments by doctors are time-consuming and costly.
• Developing automated no-reference image quality assessment (No-reference IQA) models is crucial for clinical practice.

Purpose of the Study:

• To propose a novel deep learning framework, ClinReadNet, for no-reference image quality assessment of abdominal CT scans.
• To develop a model that mimics radiologists' clinical reading logic for evaluating image quality.
• To achieve state-of-the-art performance in low-dose CT image quality assessment.

Main Methods:

• Introduced the Sobel ordinal quality network (SOQN) module for edge detail and quality distribution analysis.
• Integrated the (shifted) window multi-scale temperature multi-head self-attention ((S)W-MTMSA) module for attention-based region focusing.
• Designed the hierarchical ranked probability score (HRPS) loss function for improved classification accuracy.

Main Results:

• ClinReadNet achieved state-of-the-art (SOTA) performance on the LDCTIQAG2023 dataset.
• Achieved high correlation coefficients: Pearson's (PLCC) 0.9507, Spearman's (SROCC) 0.9554, and Kendall's (KROCC) 0.8629.
• Outperformed existing methods in image quality assessment accuracy.

Conclusions:

• ClinReadNet effectively mimics radiologists' reading habits for accurate CT image quality assessment.
• The proposed framework offers a practical solution for efficient and reliable low-dose CT image evaluation.
• This deep learning approach has significant potential to enhance diagnostic workflows in abdominal CT imaging.