Effect of a learning-based approach using the echo-signals obtained by a single-unit transducer on detection of breast tumors.

Abstract

1566 Background: Breast cancer mortalities in the less developed countries are predicted to increase over the next decades mainly due to the low utilization rates of breast cancer screening programs. Additionally, studies have shown that the sensitivity and accuracy of breast self-examination in women are suboptimal. Therefore, it is important to develop a novel and affordable technology that can accurately detect small breast tumors in women. We have developed a deep learning-based echo- signal processing system that can detect breast tumors with high sensitivity and accuracy by merely obtaining ego-motion estimated echo-signals generated by a single-unit piezoelectric transducer system. Methods: We have prospectively collected echo-signals using a single-unit piezoelectric transducer system in 200 patients with breast tumors who underwent curative surgery at the Seoul National University Hospital and SMG-SNU Boramae Medical Center between Oct 2022 and Feb 2024. The data from the first 131 cases was used for the training and testing of the deep learning model with the ratio of 8:2, and the subsequent data from 69 cases was used for the external validation. We used ImageNet pre-trained ResNet50 model to train the deep learning model and adopted various data augmentation methods for model robustness. Five nested cross-validation was used to ensure the generalization performance. Results: The median age of the enrolled patients was 55 (IOR, 47-63) and 65 patients (32.5%) had non-palpable breast tumors. Nearly half of the patients (n=93, 46.5%) had tumors smaller than 2.0 cm in the largest dimension. Our deep learning algorithm showed outstanding performance with the area under the curve (AUC) of 0.914±0.022 and 0.964±0.011 for the image- and region-level, respectively, in the training and testing set (n=131). In the external validation set (n=69), the model demonstrated an AUC of 0.846±0.012 and 0.932±0.009 for the image- and region-level, respectively. As shown (Table), the accuracy, sensitivity, and specificity were substantially improved for the region-level analysis when compared with those of the image-level analysis. Conclusions: Using a prospectively collected data from 200 patients, we developed a highly accurate and robust deep learning model to detect breast tumors in women by processing echo-signals obtained by a single-unit piezoelectric transducer system. Our data indicate that deep learning approach based on low-volume echo signal can lead to a novel and affordable breast tumor screening strategy that can improve early breast cancer detection.[Table: see text]