Multi-center Study on Artificial Intelligence-based Classification of Breast Lesions

In the field of breast cancer, early diagnosis is crucial for improving treatment efficacy and survival rate. Breast cancer can be mainly divided into two categories: in situ carcinoma and invasive carcinoma, which have significant differences in treatment strategies and prognosis. The incidence of axillary lymph node metastasis in in situ carcinoma is relatively low (1-2%), and sentinel lymph node biopsy (SLNB) is not recommended; whereas for invasive carcinoma, SLNB or axillary lymph node dissection (ALND) is necessary. Therefore, it is particularly important to accurately distinguish benign, malignant, in situ, and invasive lesions preoperatively.

Contrast-enhanced mammography (CEM) is an emerging technique that has been increasingly used in clinical practice due to its ability to reveal the vascular characteristics of lesions. However, although CEM has high sensitivity for malignant lesions in diagnosing breast cancer, its specificity is not entirely satisfactory (66-84%). Additionally, the interpretation of traditional imaging examinations is influenced by the experience of radiologists, with significant differences among radiologists. Therefore, it is necessary to develop an automated, reliable, and non-invasive method that can preoperatively distinguish between benign and malignant breast lesions, as well as between in situ and invasive carcinomas.

Deep learning is a powerful artificial intelligence (AI) technique that has gained widespread attention due to its outstanding performance in image recognition tasks. Although previous studies have applied deep learning to CEM images to predict benign and malignant breast lesions, these studies had relatively small sample sizes and lacked multi-center data validation of their generalization capabilities. Furthermore, the value of applying deep learning to distinguish between in situ and invasive carcinomas remains unclear.

Research Sources

This study was conducted jointly by researchers from multiple hospitals and research institutions in Shandong, Guangdong, Shanghai, and Beijing, including Yantai Yuzhaodin Hospital, Shandong University of Business Studies, Weifang Chinese Medicine Hospital, Sun Yat-sen Memorial Hospital, Fudan University Cancer Hospital, and Beijing Cancer Hospital. The main authors include Haicheng Zhang, Fan Lin, Tiantian Zheng, Jing Gao, and others. This paper was published in the International Journal of Surgery on January 18, 2024.

Research Process

Study Subjects and Dataset

This study included 1,430 eligible patients who underwent CEM examinations between June 2017 and July 2022. The dataset was divided into a training set (n=1,101), an internal test set (n=196), and an external test set (n=133). The study subjects were female patients who underwent CEM examinations and had histologically confirmed breast lesions.

Image Processing and Feature Extraction

In this study, the AI model used RefineNet as the backbone network, and a convolutional block attention module (CBAM) sub-network was constructed on top of the backbone network for adaptive feature optimization. The output of CBAM was applied to a global average pooling (GAP) layer to generate optimized deep learning features for CEM images.

Classification Module

The classification module utilized the XGBoost classifier in combination with optimized deep learning features and clinical features for collaborative decision-making in the preoperative diagnosis of benign and malignant breast lesions. Both RefineNet and ResNet were used as feature extraction modules for the model, with the former yielding the best results.

Reader Study

Radiologists were asked to independently assess the benignity and malignancy of breast lesions, and then reassess them with the assistance of the AI model. The results obtained by radiologists alone and with the aid of the AI model were compared to evaluate the improvement in diagnostic ability provided by the AI model.

Biological Basis Exploration

To unravel the biological basis underlying the AI predictions, gene analysis was performed based on RNA sequencing data from 12 patients. The 12 patients were grouped according to the AI-predicted high-risk and low-risk categories, and differential gene expression analysis and enrichment pathway identification were conducted using the R packages DESeq2 and ClusterProfiler.

Research Results

Main Results

In distinguishing between benign and malignant breast lesions, the AI model achieved an area under the curve (AUC) of 0.932 on the external test set, outperforming the best-performing deep learning model, radiomics model, and radiologists. Furthermore, the AI model also achieved satisfactory results in the diagnosis of in situ and invasive carcinomas (AUC ranging from 0.788 to 0.824).

Biological Basis

Gene expression analysis of the high-risk and low-risk groups revealed that the high-risk group was associated with pathways such as extracellular matrix (ECM) remodeling. These findings provided biological insights supporting the AI model’s predictions.

Conclusion

The AI model based on CEM and clinical features demonstrated excellent predictive performance in the diagnosis of breast lesions, effectively distinguishing between benign and malignant lesions, as well as further differentiating in situ carcinomas from invasive carcinomas. This study is the first to combine CEM images with AI classification and explore the biological basis of the AI model, providing valuable support for future clinical decision-making.

Research Highlights

1. The AI model, combining attention-based deep learning features and clinical features, achieved outstanding performance in the diagnosis of benign and malignant breast lesions.
2. Effective identification of in situ and invasive breast carcinomas was achieved.
3. The biological basis underlying the AI model was explored in-depth, revealing relevant gene pathways.