Breast Cancer Detection on Dual-View Sonography via Data-Centric Deep Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate breast cancer diagnosis is crucial for effective treatment.
• AI models show promise in improving diagnostic accuracy.
• Dual-view ultrasound offers comprehensive imaging data.

Purpose of the Study:

• To enhance AI-assisted breast cancer diagnosis using dual-view sonography.
• To develop and optimize a data-centric AI model for differentiating malignant and benign breast masses.
• To evaluate the performance of the AI model against radiologists and assess its assistive capabilities.

Main Methods:

• A DenseNet-based AI model was customized using a proprietary dual-view breast ultrasound dataset.
• Various strategies were explored to integrate dual-view images into the model.
• The AI model's performance was compared to that of a radiologist, and the impact of AI assistance on radiologist performance was quantified.

Main Results:

• The optimal model configuration utilized channel-wise stacking of dual-view images, achieving an AUC of 0.9754.
• The AI model demonstrated superior performance compared to the radiologist, particularly in specificity (0.96 vs. 0.48).
• AI assistance improved the radiologist's accuracy by 17%, precision by 26%, and specificity by 29%.

Conclusions:

• The customized dual-view AI model significantly outperforms existing methods and radiologists in breast cancer diagnosis.
• The AI model can serve as a valuable standalone tool or assistive aid, enhancing specificity and reducing unnecessary biopsies.
• Implementing this AI solution can alleviate radiologist workload and improve patient outcomes through more accurate diagnoses.