Tree-structured vector quantization of CT chest scans: image quality and diagnostic accuracy

Area of Science:

• Medical Imaging
• Data Compression
• Radiology

Background:

• Lossy compression algorithms are increasingly applied to medical images to reduce storage and transmission burdens.
• Previous studies often used simplified tasks or statistical methods not fully representative of clinical practice.
• Evaluating the impact of compression on diagnostic performance requires methods that mimic real-world radiologist interpretation.

Purpose of the Study:

• To assess the diagnostic performance of radiologists on lossy compressed medical images.
• To introduce and validate novel methods for quantifying image quality in the context of diagnostic tasks.
• To compare a novel compression technique (tree-structured vector quantization) with traditional methods.

Main Methods:

• Application of a low-complexity predictive tree-structured vector quantization algorithm for lossy compression of computerized tomography (CT) chest scans.
• Radiologists evaluated both uncompressed and compressed images for diagnostic tasks, including nodule and lymphadenopathy detection.
• Image quality was quantified using diagnostic performance, signal-to-noise ratios, subjective ratings, and advanced statistical methods for nonbinary data.

Main Results:

• Significant compression of CT images from 12 bits per pixel (bpp) down to 1-2 bpp was achieved.
• No significant changes in diagnostic accuracy were observed for the evaluated compression levels and tasks.
• The proposed image quality evaluation techniques proved effective and independent of the specific compression algorithm used.

Conclusions:

• Lossy compression to 1-2 bpp is feasible for CT chest scans without compromising diagnostic accuracy.
• The developed evaluation framework offers a robust method for assessing the impact of lossy image processing techniques.
• This approach supports the efficient management of medical imaging data while maintaining diagnostic integrity.