Combining Radiomics and Autoencoders to Distinguish Benign and Malignant Breast Tumors on US Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Ultrasound (US) is a key breast imaging tool, but its accuracy relies heavily on operator expertise.
• Computer-assisted real-time image analysis offers a solution to enhance diagnostic performance and overcome operator dependency.

Purpose of the Study:

• To create a precise, real-time capable US-based system for breast tumor categorization.
• To integrate classic radiomics and autoencoder-based features from automatically localized lesions for improved classification.

Main Methods:

• Retrospective analysis of 1619 B-mode US breast tumor images.
• Utilized nnU-Net for precise lesion segmentation.
• Extracted and combined classic radiomics and autoencoder features from segmented lesions and bounding boxes.
• Trained machine learning algorithms for tumor categorization, evaluated using AUC, sensitivity, and specificity.

Main Results:

• nnU-Net demonstrated high precision and reproducibility in lesion segmentation (median Dice Score 0.90).
• The best model, using 23 mixed features, achieved an AUC of 0.90, 81% sensitivity, and 87% specificity.
• Model performance showed no significant difference compared to human readers or histopathological diagnosis.

Conclusions:

• A precise real-time US-based breast tumor categorization system was successfully developed.
• The system effectively combines classic radiomics and autoencoder-based features from tumor bounding boxes.
• This AI-driven approach holds potential for improving diagnostic accuracy in breast imaging.