Radiomics analysis combining unsupervised learning and handcrafted features: A multiple-disease study.

Area of Science:

• Radiomics
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Radiomics extracts quantitative features from medical images.
• Machine learning models are increasingly used for disease identification.
• Combining different feature types may enhance diagnostic performance.

Purpose of the Study:

• To investigate the synergistic benefits of combining handcrafted and learning-based features for disease identification.
• To evaluate this hybrid approach across diverse clinical scenarios.

Main Methods:

• Retrospective analysis of 170-209 patients across four disease types (gastric cancer, osteosarcoma, cholangiocarcinoma, pancreatic neuroendocrine tumors).
• Extraction of 67 handcrafted features and sparse autoencoder (SAE)-based learning features from CT and MRI scans.
• Comparison of prediction models using handcrafted features alone, SAE features alone, and hybrid features.

Main Results:

• Hybrid features yielded the best performance in three out of four studies, with AUCs reaching 0.829 (GC), 0.758 (ICC), and 0.771 (pNETs).
• SAE features alone performed best in the high-grade osteosarcoma (HOS) study (AUC 0.740).
• Correlation analysis assessed the complementarity between handcrafted and SAE features.

Conclusions:

• Combining handcrafted and learning-based features offers general benefits in radiomics modeling for disease identification.
• Performance gains are dependent on the specific task and data.
• Study-specific feature selection and model optimization are crucial for achieving high accuracy and robustness.