Development of an algorithm to automatically compress a CT image to visually lossless threshold.

Area of Science:

• Medical Imaging
• Computer Science
• Image Compression

Background:

• Visually lossless thresholds (VLTs) are crucial for optimizing CT image compression.
• Accurate VLT prediction aids in balancing image quality and file size for medical data.
• Current methods for VLT determination can be computationally intensive.

Purpose of the Study:

• To develop and evaluate a novel algorithm for predicting CT image VLTs using only original image features and DICOM header information.
• To assess the algorithm's performance against established image fidelity metrics like PSNR and HDR-VDP.
• To determine the computational efficiency of the proposed VLT prediction model.

Main Methods:

• A multiple linear regression (MLR) model was constructed using training data (n=103 CT images).
• Independent variables included image features and DICOM header data; the dependent variable was radiologist-determined VLT (VLTrad).
• Model performance was validated on a separate testing set (n=103) and compared with PSNR and HDR-VDP using absolute differences and intra-class correlation (ICC).

Main Results:

• The MLR model achieved a mean absolute difference of 0.58 from VLTrad, outperforming PSNR (0.73) and HDR-VDP (0.68) (p<0.01).
• The MLR model demonstrated a higher ICC (0.88) compared to PSNR (0.85) and HDR-VDP (0.84).
• Computational time for the MLR model (1.5s) was significantly faster than PSNR (3.9s) and HDR-VDP (68.2s).

Conclusions:

• The proposed MLR algorithm effectively predicts CT image VLTs using image features and DICOM data.
• The model provides comparable or superior performance to existing metrics with substantially lower computational expense.
• This algorithm shows promise for enabling adaptive compression of CT images.