A Comparison of Computer-Aided Diagnosis Schemes Optimized Using Radiomics and Deep Transfer Learning Methods

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Diagnostic Systems

Background:

• Radiomics and deep transfer learning (DTL) are key technologies for developing computer-aided detection and diagnosis (CAD) schemes.
• Comparing their efficacy and limitations in CAD development is crucial for advancing medical image analysis.

Purpose of the Study:

• To investigate and compare the advantages and limitations of radiomics and DTL for developing CAD schemes.
• To evaluate the performance of two distinct CAD schemes for breast lesion classification.

Main Methods:

• A retrospective dataset of 3000 digital mammograms (1496 malignant, 1504 benign) was utilized.
• Two CAD schemes were developed: one using radiomics features and a support vector machine classifier, the other using a pre-trained ResNet50 model for DTL.
• Both schemes underwent 10-fold cross-validation, with performance evaluated using the area under the ROC curve (AUC).

Main Results:

• The DTL-based CAD scheme (ResNet50) achieved a significantly higher AUC of 0.85 ± 0.02 compared to the radiomics-based scheme (AUC = 0.77 ± 0.02).
• Score fusion of the two CAD schemes did not lead to improved classification performance.
• The ResNet50 model demonstrated superior lesion classification capability.

Conclusions:

• Deep transfer learning offers a more efficient approach for developing CAD schemes in medical imaging.
• DTL-based CAD schemes provide higher lesion classification performance than those relying solely on radiomics features.
• This study underscores the potential of DTL for enhancing diagnostic accuracy in mammography.