Recovery of GLRLM Features in Degraded Images using Deep Learning and Image Property Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiomics

Background:

• Radiomics models predict clinical outcomes but suffer from variability due to diverse imaging conditions.
• This limits their generalizability and reliability in real-world applications.

Purpose of the Study:

• To extend a dual-domain deep learning approach for recovering ground truth radiomics feature values.
• Specifically, to address the restoration of gray-level run length matrix (GLRLM) features.

Main Methods:

• Developed a novel algorithm for the differentiable approximation of GLRLMs.
• Implemented a dual-domain deep learning network with a dual-domain loss function.
• Assessed performance using lung CT image patches, focusing on feature recovery accuracy and classification performance.

Main Results:

• The proposed network achieved the lowest Mean Squared Error (MSE) in GLRLM feature recovery compared to baselines.
• A classification model using recovered GLRLM features reached 86.65% accuracy.
• This closely matched the 88.85% accuracy of models using ground truth features, outperforming degraded features (82.00%).

Conclusions:

• The developed deep learning method effectively restores GLRLM radiomics features degraded by imaging conditions.
• This approach shows significant potential for standardizing radiomics and improving the generalizability of predictive models.
• The method enhances the accuracy of clinical outcome prediction using radiomics analysis.