A Combined Segmentation and Classification Pipeline for Breast Tumors Analysis on Ultrasound Image
Cong Thanh Nguyen, Quang Linh Huynh · Journal of Physics Conference Series · 2025
Abstract Breast cancer is a life-threatening disease characterized by the abnormal and uncontrollable growth of breast cells, leading to tumor development. Ultrasound is currently an essential non-invasive imaging technique for evaluating the features of breast tumors. During the diagnostic process, radiologists perform tumor segmentation and classification into benign or malignant categories. This manual process presents challenges due to the need for high accuracy to ensure effective diagnosis. Therefore, an automated approach is necessary to enhance precise tumor segmentation and classification as a technical diagnostic tool. This study developed a tool that integrates automated segmentation and classification of breast tumor ultrasound images using deep learning models. Firstly, the tumor segmentation process was implemented with a Deep Residual UNET model to identify the suspect region on breast ultrasound images. The original breast ultrasound image was then combined with the identified tumor area from the segmentation process to increase the information available during the classification process. The VGG16 model was ultimately employed to classify breast tumors as either benign or malignant. These two deep learning models were trained on a public breast ultrasound dataset comprising 437 benign and 210 malignant tumors. Model validation was conducted using 5-fold cross-validation. The segmentation-alone model achieved an accuracy of 98.93% ± 0.40% and a Dice coefficient of 89.57% ± 2.16%. The classification model and the combined model achieved mean accuracies of 98.3% and 78%, respectively, and weighted F1-scores of 98.30% and 78.27%, respectively. This work presents a combined breast tumor segmentation and classification tool with considerable performance. Nevertheless, additional efforts are required to enhance the performance of the combined model.