Predictive model of breast cancer lymph node metastasis based on deep learning (E-Transformer)

• Published in: 2021 6th International Symposium on Computer and Information Processing Technology (ISCIPT) pp. 168 - 173
• Main Authors: Wang, Qian, Qiao, Kai, Qin, RuoXi, Hai, JinJin, Yan, Bin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2021
• Subjects: Artificial intelligence; auxiliary diagnosis; Breast cancer; Convolutional neural networks; Deep learning; E-Transformer; Feature extraction; lymphatic metastasis; mammography; medical image analysis; Metastasis; Predictive models; Transformers
• DOI: 10.1109/ISCIPT53667.2021.00041

Breast cancer cell lymph node metastasis is an important criterion to define the early and middle stages of breast cancer with biopsy required, and so with bad patient experience. Traditional image-aided diagnosis is inefficient and ineffective in manually extracting features and combining image features. The emerging image-assisted diagnosis based on deep learning uses convolutional neural networks to automatically segment lesions, extract image features, and automatically combine features to classify cancers through full-connected layers or machine learning, which provides new ideas for clinicians' diagnosis and treatment plans. However, whether cancer cells have lymph node metastasis or not has little difference in Molybdenum mammography of breast, and it is difficult to distinguish, which belongs to the problem of fine-grained image classification. If two images are randomly selected, even the most experienced doctors in the radiology department cannot directly judge whether the cancer cells have metastasized in the lymph nodes from the mammography images. This paper proposes a breast cancer lymph node metastasis prediction model based on deep learning, named E-Transformer, which solves the problem of fine-grained classification of breast cancer lymph node metastasis. The model uses EfficientNet for coarse-grained feature extraction, adds Transformer-encoder to introduce an attention mechanism to enhance the model's fine-grained feature processing capabilities and finally uses Lightgbm for feature combination and two categories. Tests on the unpublished mammography data set of 391 breast cancer patients found that the auc of this model is 0.9643, the ace is 0.9681, and the f1-score is 0.9669. The reliability and accuracy of the model are in the first echelon of the current study.