Deep learning combined with attention mechanisms to assist radiologists in enhancing breast cancer diagnosis: a study on photoacoustic imaging

Abstract

Accurate prediction of breast cancer (BC) is essential for effective treatment planning and improving patient outcomes. This study proposes a novel deep learning (DL) approach using photoacoustic (PA) imaging to enhance BC prediction accuracy. We enrolled 334 patients with breast lesions from Shenzhen People’s Hospital between January 2022 and January 2024. Our method employs a ResNet50-based model combined with attention mechanisms to analyze photoacoustic ultrasound (PA-US) images. Experiments demonstrated that the PAUS-ResAM50 model achieved superior performance, with an AUC of 0.917 (95% CI: 0.884 –0.951), sensitivity of 0.750, accuracy of 0.854, and specificity of 0.920 in the training set. In the testing set, the model maintained high performance with an AUC of 0.870 (95% CI: 0.778–0.962), sensitivity of 0.786, specificity of 0.872, and accuracy of 0.836. Our model significantly outperformed other models, including PAUS-ResNet50, BMUS-ResAM50, and BMUS-ResNet50, as validated by the DeLong test (p < 0.05 for all comparisons). Additionally, the PAUS-ResAM50 model improved radiologists’ diagnostic specificity without reducing sensitivity, highlighting its potential for clinical application. In conclusion, the PAUS-ResAM50 model demonstrates substantial promise for optimizing BC diagnosis and aiding radiologists in early detection of BC.