Superpixel assisted deep neural network for breast tumor segmentation in ultrasound images

Abstract

Breast cancer is the leading type of cancer diagnosed in women, according to data from the World Health Organization (WHO) in 2020. It is also the type of cancer that causes the most deaths in women, with around 685,000 deaths. In the diagnosis of breast cancer, ultrasound imaging has been used frequently. Tumor segmentation from breast ultrasound (BUS) images is significant for the success of subsequent analyses such as tumor detection, classification, and treatment planning in computer-aided diagnostic (CAD) systems. However, traditional segmentation approaches are challenging to apply automatically since tumor size, shape, and echo intensity vary significantly among BUS images. Deep learning-based segmentation approaches have great potential to reduce the workload of radiologists and operator dependence by automating tedious tasks. However, tumor segmentation remains a difficult task for these approaches because of the high speckle noise, artifacts, poor contrast, and intensity inhomogeneity in BUS images. Even though tumor segmentation in BUS images is challenging owing to the nature of ultrasound images, accurate lesion segmentation by reducing human involvement is essential for easier breast cancer analysis and diagnosis. For this purpose, a new superpixel-assisted deep learning model is proposed, focusing on automatic binary class breast tumor segmentation. In this thesis, a superpixel-guided deep learning network, in which residual and channel attention blocks are integrated into the U-Net network, was proposed to address the above problems. The network contains a secondary input consisting of the corresponding superpixel images in addition to the main input comprising of BUS images. Firstly, the input image is over-segmented into primitive superpixel regions with texture consistency while less semantics using the simple linear iterative clustering (SLIC) algorithm to avoid speckle noise interference while enhancing the salience of tumors in the input image. Obtaining superpixels as a pre-segmentation method and feeding it as a second input to the network will provide good guidance in tumor segmentation. On the other hand, because there are far fewer superpixels than pixels in the input image, using superpixels can greatly lower the overall computational cost. Experiments in this thesis showed that the use of superpixel images can improve the tumor segmentation success of the proposed model. Images that were utilized to train and evaluate the breast tumor segmentation algorithm developed by this thesis were taken from two publicly available BUS image datasets, BUSI and UDIAT. A number of data augmentation and normalization techniques are applied to these datasets. In proposed U-Net-based model, residual blocks were placed in both encoder path and decoder path. Residual modules improve feature extraction and expression, as well as resolve degradation issues, allowing for greater accuracy gains with higher depth. Additionally, the output of residual blocks in the encoder part is passed through the Channel Attention (CA) block to increase the network's representational power. The channel attention mechanism can explore the interdependence between the feature channels to boost segmentation performance. The Superpixel Channel Attention (SCA) module is a combination of superpixel features and weighted channel information derived by the CA block. This module is a channel attention enriched module for integrating prior knowledge of superpixel. Furthermore, bottlenecks in U-shaped convolutional neural networks are a way to force the model to learn a compression of the input data. The idea is that this compressed view should only contain the useful information to be able to construct an output mask. Therefore, because the high-level feature map represents complex features with wide receptive fields and more channels, the Channel Attention Residual (CAR) module was added to the model's bottom layer. The training was repeated with the 5-fold cross-validation technique to obtain a more consistent model for all cases. The final pixel labels are voted in the final assessment using ensemble models with random parameter initializations in 5-fold data. It turns out that a final segmentation output from an ensemble of models trained with different inputs and using the same architecture outperforms a single model. Ensemble learning application on superpixel-guided deep learning network, which is the recommended approach for breast lesion segmentation, gives better results than all competing U-Net variant models and either of the 5-fold models. Test results within the scope of the thesis showed that tumor segmentation from breast ultrasound images can be effectively accomplished using a method that combines deep neural networks and superpixel information.