Application of dual-modality ultrasound deep learning predictive model in diagnosing breast cancer

Abstract

Objective To develop a predictive model using deep learning (DL) techniques based on breast ultrasound grayscale images and elastography, and to explore the diagnostic efficacy of this model in differentiating benign from malignant breast lesions. Methods Data of 1 000 breast lesions from patients who underwent surgical treatment at Linyi People's Hospital Breast Surgery Department from May 2020 to April 2021 were collected retrospectively. including ultrasound images and related clinical pathological information. The ultrasound grayscale image and elastography of the largest section of each lesion were selected and randomly divided into training, validation, and test sets at a ratio of 7∶2∶1. A predictive model was constructed based on neural networks using the training and validation sets, and the diagnostic efficacy of the model was tested with the test set images. Four sonographers were invited to read the test set ultrasound images independently, and their diagnostic efficacies were compared with the model's performance. Results The area under the curve (AUC) value (0.907) of the receiver operating characteristic curve (ROC) of DL model for breast lesion diagnosis was higher than all participating sonographers, with a statistically significant difference (P＜0.05). The average AUC value for the diagnosis by senior sonographers (0.835) was higher than that for junior sonographers (0.719), with a statistically significant difference (P＜0.05). When the model assisted junior sonographers and senior sonographers in diagnosing the test set breast lesions, the average AUC value was 0.806 and 0.864, respectively.〖LM〗 After assistance from the model, the diagnostic efficacy of sonographers of different experience levels improved, with a more significant increase for junior sonographers (P＜0.05). Notably, there was no statistically significant difference in the AUC values between the junior sonographers assisted by the DL model and the senior sonographers reading alone (P＞0.05). Conclusion A predictive model based on dual-modality ultrasound DL can significantly improve the diagnostic efficacy of sonographers in differentiating benign from malignant breast lesions.

Keywords

• deep learning
• neural network
• breast cancer
• ultrasound
• elastic imaging
• grayscale image
• prediction model
• artificial intelligence