Automatic Convolutional Neural Network Architecture Search for Breast Lesion Classification from Ultrasound Images: An ENAS Bayesian Optimization Approach

Ahmed, Mohammed Hussein (2022) Automatic Convolutional Neural Network Architecture Search for Breast Lesion Classification from Ultrasound Images: An ENAS Bayesian Optimization Approach. Doctoral thesis, The University Of Buckingham.

Text Ahmed Mohammed 1200526 Thesis.pdf - Submitted Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB)

Abstract

Breast cancer is one of the most common cancer types among women globally. Cancer detection/classification from medical images is a topic of vital importance because early cancer detection may allow patients to receive proper and timely treatment, significantly increasing their survival rates. Ultrasound imaging has been extensively used for various medical diagnostics by radiologists. Over the last ten years, sophisticated deep learning neural networks such as Convolutional Neural Networks (CNN) have been developed. Such deep learning neural networks tend to provide an “end-to-end” solution for image pattern recognition and have achieved impressive performance results for various applications. Deep convolutional neural networks have recently appeared in CAD systems due to their success in extracting effective image features. CNN architecture design involves using many hyper-parameters. Creating a robust CNN architecture depends on finding an optimal combination between those hyper-parameters. Therefore, manually designing CNN architecture is time-consuming and leads to trial and error. Neural Architecture Search offers an alternative by automatically determining hyper-parameter settings for CNN architectures based on the dataset at hand. Efficient Neural Architecture Search (ENAS) is the efficient method for automatically designing CNN architecture. There has been no research in the literature - till that reported in this thesis - on using ENAS to search for CNN architecture for ultrasound images in general nor for breast lesion classification. In addition, there are only a few reported pieces of research on CNN architectures manually designed specifically for breast lesion classification from ultrasound images. These research works are based on only small datasets from one hospital in their modelling and testing processes. This research aims to create an automatic designing CNN architecture approach for designing CNN architectures for classifying breast cancer from ultrasound images. This research investigates the effectiveness of one of the most popular methods, ENAS, for automatically searching for a convolutional neural network architecture. The research starts by adapting the ENAS framework to automatically search for optimal CNN architectures based on datasets of ultrasound images collected from different medical centres. The research then addresses the issue of model overfitting and generalisation of ENAS-based CNNs by using different data augmentation, reducing architecture complexity and training on an unbalanced number of images between benign and malignant classes. Furthermore, the ENAS framework is modified by expanding its search space by adding more operations suitable for ultrasound images such as different convolutional operations with different filter sizes. This modification improved the overall performance of produced CNN architecture by ENAS. We further enhanced the design of final CNN model which are based on optimal cells obtained by ENAS by adding a high-way connection to compensate features from early layers to the final set of feature maps. Furthermore, this research deploys the Bayesian Optimisation method to further develop an ENAS-B framework to address the limitations of the existing ENAS framework in optimizing the CNN architecture layers and trainable hyper-parameters, promoting an end-to-end automatic CNN search for the intended purpose of breast lesion classification using ultrasound images. The research concludes that a CNN model of 5 layers with optimised hyper-parameters is a robust model that can outperform the state-of-the-art CNN models designed for breast lesion classification such as VGG-16, ResNet50, Inception-V3, XceptionNet, NasNet mobile, EfficentNetB0 and mobile Net-V2 methods with transfer learning. The work presented in this thesis provides a good guideline for scientists to design a robust CNN model that can generalise beyond internal testing datasets.

Actions (login required)

View Item