BoostCNN: Deep Learning AdaBoost-based Method for Easy and Difficult Nodule Classification in Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast ultrasonography complements mammography but faces challenges with high false positive rates.
• Lack of consensus exists on ultrasound's primary role in breast cancer screening due to over-diagnosis concerns.
• False positives in breast ultrasound can lead to unnecessary procedures and patient distress.

Purpose of the Study:

• To develop a novel AdaBoost-based ensemble method (BoostCNN) for breast cancer detection using ultrasound.
• To reduce false positives and false negatives in breast cancer ultrasound classification.
• To enhance the specificity and accuracy of automated breast ultrasound analysis.

Main Methods:

• Implemented and evaluated a novel AdaBoost-based ensemble deep learning model, BoostCNN.
• Trained and tested models on the BUS-BRA dataset, including state-of-the-art deep learning approaches.
• Performed external validation on the BUSI and BrEaSt datasets to assess generalizability.

Main Results:

• The BoostCNN ensemble with nine models achieved 88.38% accuracy, 82.90% sensitivity, and 91.00% specificity on the BUS-BRA dataset.
• Combined ultrasonographer assessment with BoostCNN prediction improved performance, yielding 88.21% accuracy and 90.13% specificity.
• The proposed method demonstrates potential for enhancing specificity in breast cancer ultrasound classification.

Conclusions:

• BoostCNN offers a promising approach to improve the accuracy of breast cancer detection via ultrasound.
• The ensemble method effectively reduces false positives and negatives, addressing key limitations of current ultrasound screening.
• This AI tool has the potential to support specialists, leading to more reliable breast cancer diagnosis.