Generative models improve radiomics performance in different tasks and different datasets: An experimental study.

Area of Science:

• Medical imaging analysis
• Radiomics
• Deep learning

Background:

• Radiomics extracts quantitative features from medical images for clinical decision support, particularly in oncology.
• Noise in low-dose computed tomography (CT) scans hinders accurate radiomic feature extraction.
• Deep learning generative models offer a potential solution for noise reduction in low-dose CT images.

Purpose of the Study:

• To investigate the use of deep learning generative models to improve radiomics performance on low-dose CT scans.
• To assess the impact of denoising on survival prediction and lung cancer diagnosis using radiomic features.

Main Methods:

• Utilized encoder-decoder networks and conditional generative adversarial networks (CGANs) to transform low-dose CT images into full-dose images.
• Extracted radiomic features from both original and denoised low-dose CT scans.
• Developed support vector machine (SVM) and deep attention-based multiple instance learning models for classification tasks.

Main Results:

• Denoising improved the area under the curve (AUC) for survival prediction from 0.52 to 0.57 (p < 0.01).
• AUC for lung cancer diagnosis increased from 0.84 to 0.88 (encoder-decoder) and 0.89 (CGAN) (p < 0.01).
• No significant AUC improvement was observed when models were trained at 75 and 100 epochs.

Conclusions:

• Generative models effectively enhance radiomics performance for tasks utilizing low-dose CT data.
• Denoising low-dose CT scans with generative models is a vital preprocessing step for accurate radiomic feature calculation.
• This approach shows promise for improving clinical decision support in oncology.