Ultrasound Image Classification and Segmentation Using Deep Learning Applications PDF Download

Author: Umar Farooq Mohammad
Publisher:
ISBN:
Category :
Languages : en
Pages : 0

Get Book

Book Description
Breast cancer is one of the most common diseases with a high mortality rate. Early detection and diagnosis using computer-aided methods is considered one of the most efficient ways to control the mortality rate. Different types of classical methods were applied to segment the region of interest from breast ultrasound images. In recent years, Deep learning (DL) based implementations achieved state-of-the-art results for various diseases in both accuracy and inference speed on large datasets. We propose two different supervised learning-based approaches with adaptive optimization methods to segment breast cancer tumours from ultrasound images. The first approach is to switch from Adam to Stochastic Gradient Descent (SGD) in between the training process. The second approach is to employ an adaptive learning rate technique to achieve a rapid training process with element-wise scaling in terms of learning rates. We have implemented our algorithms on four state-of-the-art architectures like AlexNet, VGG19, Resnet50, U-Net++ for the segmentation task of the cancer lesion in the breast ultrasound images and evaluate the Intersection Over Union (IOU) of the four aforementioned architectures using the following methods : 1) without any change, i.e., SGD optimizer, 2) with the substitution of Adam with SGD after three quarters of the total epochs and 3) with adaptive optimization technique. Despite superior training performances of recent DL-based applications on medical ultrasound images, most of the models lacked generalization and could not achieve higher accuracy on new datasets. To overcome the generalization problem, we introduce semi-supervised learning methods using transformers, which are designed for sequence-to-sequence prediction. Transformers have recently emerged as a viable alternative to natural global self-attention processes. However, due to a lack of low-level information, they may have limited translation abilities. To overcome this problem, we created a network that takes advantages of both transformers and UNet++ architectures. Transformers uses a tokenized picture patch as the input sequence for extracting global contexts from a Convolution Neural Network (CNN) feature map. To achieve exact localization, the decoder upsamples the encoded features, which are subsequently integrated with the high-resolution CNN feature maps. As an extension of our implementation, we have also employed the adaptive optimization approach on this architecture to enhance the capabilities of segmenting the breast cancer tumours from ultrasound images. The proposed method achieved better outcomes in comparison to the supervised learning based image segmentation algorithms.