Che, Hui. Improved nonalcoholic fatty liver disease diagnosis from ultrasound data based on deep learning. Retrieved from https://doi.org/doi:10.7282/t3-wwwc-hp57
DescriptionNonalcoholic fatty liver disease (NAFLD) is one of the most common chronic liver diseases and its incidence increases year by year on the global scale as the quality of life improves. The asymptomatic characteristics of NAFLD prevents timely diagnosis, so it is becoming more and more urgent to find effective diagnosis and treatment methods to serve patients. Ultrasound (US) is the preferred noninvasive modality to detect steatosis due to its real-time and cost-effective imaging capabilities. However, the sensitivity of traditional B-mode US cannot support steatosis detection, which has a fat content less than 20%. Furthermore, current practice mostly relies on visual investigation of the collected data by expert clinicians there is qualitative and very subjective. With the success of deep learning-based methods applied in medical image analysis, various convolutional neural networks (CNNs) draw a lot of attention and have been investigated for classifying liver diseases from US data as well. The purpose of the deep learning is to automatically learn high-level abstractions from medical data. However, designing accurate and robust CNN architectures requires large amount of input data and balanced classes. These issues remain to be significant challenges in medical image processing since the available medical data is scarce. Therefore, many studies consider to overcome these shortcomings at the same time for facilitating model classification performance.
In this thesis, novel deep learning models are proposed for NAFLD classification and image synthesis from B-mode US data. For the main task of classification, we design a multi-feature guided multi-scale residual CNN architecture to capture features of different receptive fields. Besides the use of original data, an image processing method for feature enhancement is introduced for diverse and robust information. B-mode US images are combined with their corresponding local phase filtered images and radial symmetry transformed images as multi-feature inputs for the network. Various fusion strategies are studied to improve prediction accuracy. To address the problems of small datasets and class imbalance, a computational method based on Generative Adversarial Network (GAN) is proposed to synthesize B-mode US images. In order to make the generated images look more realistic, self-attention blocks are added to the network for the maintenance of important structural information. We evaluate the designed networks on 550 B-mode in vivo liver US images collected from 55 subjects. Quantitative results show an average classification accuracy above 90% over 10-fold cross-validation. Comprehensive validation results further demonstrate that our proposed approaches achieve significant improvements compared to the state-of-the-art medical image classification methods. Furthermore, the image synthesis network is validated using the whole image and cropped patches, and results show it has a strong ability to generate the whole image with realistic details. In the situation that a larger balanced NAFLD training dataset is produced including real and synthesized data, the multi-feature multi-scale CNN acquires better classification performance no matter according to the result from the whole image or patch data. Based on the promising performance, the proposed methods have the potential in practical applications to help radiologists diagnose NAFLD.