Segmenting Abnormalities and Predicting Breast Cancer in Ultrasound Images using Modified UNET Architecture

Abstract

Breast cancer is still a major worldwide health concern, and better patient outcomes and effective treatment depend on early identification. Different algorithms try to classify the breast cancer either malignant or benign or try to segment the abnormal section with the medical images. Convoluted Networks like Convolutional Neural Network (CNN) are broadly used for classification whereas U-shaped network like UNet, DeepLab are used for segmentation. This study suggests a multi-tasking UNet architecture where a single model perform both the classification and segmentation task over breast cancer BUSI dataset. Two different nature of dataset a) grayscale USG image with labels and b) grayscale with ground truth mask is sent as input. The model is trained under the train test split ratio of 80:10:10. In classification, the model achieved 98.360.62% of Training accuracy, 98.30 ± 0.94 % of Validation accuracy and 98.08 ± 0.64% of Testing Accuracy along with 0.19 ± 0.31, 0.092 ± 0.13 and 0.122 ± 0.18 Training loss, Validation loss and Testing loss respectively, whereas in Segmentation, the model achieved Intersection over Union (IoU) value of is 89.089%. The achieved results hold significant promise for advancing the field of medical image analysis, ultimately contributing to improved diagnosis and treatment outcomes for breast cancer patients.