Biomarker Localization From Deep Learning Regression Networks.

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Biomarker Quantification

Background:

• Traditional biomarker estimation relies on segmentation-then-measurement, requiring segmentation masks.
• Deep learning regression networks estimate biomarkers directly but function as black boxes.
• Existing methods lack simultaneous quantitative results and qualitative localization assessment.

Purpose of the Study:

• To develop a novel deep learning network for simultaneous biomarker regression and localization mask generation.
• To train the network using only the biomarker value, eliminating the need for segmentation masks.
• To evaluate the proposed method's performance against existing direct regression and segmentation-based approaches.

Main Methods:

• Proposed a novel deep learning network structure capable of joint biomarker regression and localization.
• Trained and evaluated three network variations on four medical imaging problems: pectoralis muscle area (PMA), subcutaneous fat area (SFA), liver mass area (CT), and coronary artery calcium (CAC).
• Compared performance against direct regression networks and U-Net segmentation-based quantification.

Main Results:

• The proposed method demonstrated improved performance over direct regression methods, achieving higher correlation coefficients (e.g., 0.978 vs. 0.971 for PMA).
• Achieved significant localization accuracy with DICE coefficients of 0.875 for PMA and 0.914 for SFA.
• Outperformed U-Net segmentation-based quantification in regression accuracy for PMA and SFA.

Conclusions:

• It is feasible to achieve accurate simultaneous biomarker regression and localization using deep learning trained solely on biomarker values.
• The proposed method offers a more interpretable and potentially more accurate alternative to black-box regression and traditional segmentation approaches.
• This technique enhances the quantitative and qualitative assessment of biomarkers in medical imaging.