Deep Learning-Based Computer-Aided Diagnosis for Breast Lesion Classification on Ultrasound: A Prospective...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Computer-aided diagnosis (CAD) systems for breast ultrasound are typically evaluated by expert radiologists in urban settings.
• The effectiveness of CAD in non-expert hands at rural or secondary hospitals remains less understood.

Purpose of the Study:

• To assess the utility of deep learning-based CAD software in improving diagnostic performance for radiologists lacking specialized breast ultrasound expertise.
• To evaluate CAD's role in differentiating benign from malignant breast lesions up to 2.0 cm in secondary or rural hospital settings.

Main Methods:

• Prospective study involving 313 patients across eight Chinese secondary/rural hospitals (Nov 2021-Sep 2022).
• Non-expert radiologists interpreted breast lesions (BI-RADS category 3-5) with and without CAD assistance.
• CAD was used to upgrade BI-RADS 3 to 4A and downgrade BI-RADS 4A to 3; histology confirmed diagnoses.

Main Results:

• CAD application significantly improved diagnostic accuracy (86.6% vs 62.6%), specificity (82.9% vs 46.0%), and positive predictive value (72.7% vs 46.5%).
• CAD upgraded 6.0% of BI-RADS 3 lesions to 4A (16.7% malignant) and downgraded 79.1% of BI-RADS 4A lesions to 3 (4.6% malignant).
• Sensitivity and negative predictive value showed no significant difference with CAD use.

Conclusions:

• Deep learning-based CAD software significantly enhances diagnostic performance for non-expert radiologists interpreting breast ultrasounds.
• CAD shows potential to decrease the rate of benign breast biopsies, particularly in resource-limited settings.
• The findings support CAD's role in improving breast cancer diagnosis accessibility and patient care.