Predicting Very Early-Stage Breast Cancer in BI-RADS 3 Lesions of Large Population with Deep Learning

TL;DR

This study uses deep learning to improve the detection of very early-stage breast cancer in BI-RADS 3 lesions, outperforming radiologists in accuracy.

Contribution

A novel transfer learning-based deep learning model is developed to detect early-stage breast cancer in BI-RADS 3 lesions using ultrasound images.

Findings

The model achieved an AUC of 0.880 in internal testing and 0.910 in external testing for predicting malignancy.

The model outperformed six radiologists in the external testing set with an average AUC of 0.653.

The proposed method improved the clinical AUC from 0.721 to 0.880 for predicting BI-RADS 3 malignancy.

Abstract

Breast cancer accounts for one in four new malignant tumors in women, and misdiagnosis can lead to severe consequences, including delayed treatment. Among patients classified with a BI-RADS 3 rating, the risk of very early-stage malignancy remains over 2%. However, due to the benign imaging characteristics of these lesions, radiologists often recommend follow-up rather than immediate biopsy, potentially missing critical early interventions. This study aims to develop a deep learning (DL) model to accurately identify very early-stage malignancies in BI-RADS 3 lesions using ultrasound (US) images, thereby improving diagnostic precision and clinical decision-making. A total of 852 lesions (256 malignant and 596 benign) from 685 patients who underwent biopsies or 3-year follow-up were collected by Southwest Hospital (SW) and Tangshan People’s Hospital (TS) to develop and validate a deep learning model based on a novel transfer learning method. To further evaluate the performance of the model, six radiologists independently reviewed the external testing set on a web-based rating platform. The proposed model achieved an area under the receiver operating characteristic curve (AUC), sensitivity, and specificity of 0.880, 0.786, and 0.833 in predicting BI-RADS 3 malignant lesions in the internal testing set. The proposed transfer learning method improves the clinical AUC of predicting BI-RADS 3 malignancy from 0.721 to 0.880. In the external testing set, the model achieved AUC, sensitivity, and specificity of 0.910, 0.875, and 0.786 and outperformed the radiologists with an average AUC of 0.653 (p = 0.021). The DL model could detect very early-stage malignancy of BI-RADS 3 lesions in US images and had higher diagnostic capability compared with experienced radiologists.

Figures