Breast nodule classification with two-dimensional ultrasound using Mask-RCNN ensemble aggregation

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Oncology

Background:

• Accurate differentiation of benign and malignant breast nodules is crucial for patient management.
• BI-RADS 3 and 4 classifications represent indeterminate lesions requiring further evaluation.
• Current diagnostic methods can be invasive or lack definitive accuracy.

Purpose of the Study:

• To develop a deep learning algorithm for classifying breast nodules as benign or malignant.
• To utilize two-dimensional (2D) B-mode ultrasound data for nodule characterization.
• To specifically target nodules initially categorized as BI-RADS 3 and 4.

Main Methods:

• An ensemble of mask region-based convolutional neural networks (Mask-RCNN) was employed.
• The algorithm performed both nodule segmentation and classification.
• Probabilities of malignancy were aggregated for final inference, with Area Under the Curve (AUC) used for assessment.

Main Results:

• The study included 460 breast nodule ultrasound images (295 benign, 165 malignant) for training and validation.
• A separate test set of 137 breast nodules was used, with unknown benign/malignant distribution.
• The algorithm achieved an AUC of 0.69 on the training set and 0.67 on the test set.

Conclusions:

• The proposed deep learning algorithm effectively classifies benign and malignant breast nodules.
• The system relies solely on 2D ultrasound images, simplifying the diagnostic workflow.
• This AI-driven approach offers a non-invasive method for characterizing indeterminate breast lesions.