Improving Breast Cancer Diagnosis through Multi-Class Segmentation using Attention UNet Model

One of the prominent causes of death for women universally is breast cancer. The number of early deaths is reduced with early detection. Using an ultrasound scan, the data examine the ultrasound images of breast cancer. Three categories, including normal, benign, and malignant images, comprise the Breast Ultrasound Dataset. Incorporating machine learning with breast ultrasound images might yield significant outcomes for the identification, classification, and segmentation of breast cancer. Through a mediocre image size of 500 500 pixels, the collection comprises 750 images. Attention U-Net/U- Net is the best-known model for Multi-Class Segmentation to create an attention U-Net Model. The Model will consist of a Decoder Block Encoder Block and an Attention Gate. The attention gate layers have neurons with sizes of 256, 128, and 64, which typically correspond to the number of channels in the feature maps. The output layer in the U-Net model characteristically consists of a single convolutional layer through a kernel size of 1 1. The activation function used in the output layer was sigmoid; whichever is collective for binary classification tasks like image segmentation. The sigmoid activation function outputs values between 0 and 1, representing the possibility of each pixel fitting to the target class. The loss function stated is “binary_crossentropy,” which is appropriate in binary classification tasks. The optimizer specified has Adam, which is well-liked for neural network training because of its adaptable learning rate characteristics. In contrast to conventional gradient descent techniques, the Adam optimizer adjusts the learning rate during training, which might result in faster convergence and improved results with 98% accuracy at epoch 17 with a validation loss of 0.05.