Lim, Wei Xiang
(2024)
Diabetic retinopathy image classification with quantitative saliency-oriented visual explanation and SLIC-G image preprocessing methods.
PhD thesis, University of Nottingham Malaysia.
Abstract
The field of medical image analysis has seen remarkable progress, primarily propelled by advancements in Convolutional Neural Networks (CNNs). These sophisticated neural networks have demonstrated their value across various applications, including the detection of lung cancer and segmentation of brain tumors. Such instances highlight CNNs’ potential to significantly enhance the accuracy of medical image analysis, suggesting their role in improving clinical outcomes. Despite these advancements, the inherently opaque nature of CNNs often categorizes them as ”black boxes”, raising concerns about their transparency.
Explainable Artificial Intelligence (XAI) represents a crucial approach to address the transparency concerns associated with Convolutional Neural Networks (CNNs) by elucidating the decision-making processes within these complex computational models. Gradient-based visual explanation methods have emerged as the most prevalent XAI techniques in medical image analysis. However, the validity of these gradient-based methods—where validity refers to the assessment of the explanation’s accuracy provided by the method and its alignment with the expectations of end-users remains in question. To address this gap, the first contribution of this research focused on quantitatively validating the heatmaps generated by gradient-based visual explanations, producing empirical scores that attest to their reliability. This validation is paramount for ensuring that the interpretations offered by XAI methods are not only technically accurate but also meaningful and trustworthy to end-users, including medical professionals.
Building upon the foundation of enhancing interpretability through valid explanations, the research recognizes that the performance ofCNNs in medical image analysis is also contingent upon the quality of input data, necessitated by effective image preprocessing. Image characteristics and quality significantly impact algorithm performance and prediction accuracy. Traditional preprocessing methods, while widely used, fall short in addressing the pixel-related challenges inherent in digital image processing, such as the discretization of image information and the computational burdens imposed by large images with high pixel counts. Bridging this gap, the second contribution introduces a novel superpixel-based image preprocessing method. This method not only mitigated computational demands but also optimized the image data for improved classification outcomes. In doing so, it complemented the first contribution by ensuring that the CNNs are operating on high-quality, efficiently processed images, thereby maximizing the utility and applicability of the validated explainable AI techniques. Together, these contributions represent a comprehensive approach to enhancing the transparency, interpretability, and efficiency of CNNs in medical image analysis, addressing both the 'why' and the 'how' of their decision-making processes.
In conclusion, the initial contribution demonstrated that the proposed quantitative evaluation method, focused on saliency evaluation, successfully quantitatively assess the created heatmaps by generating empirical scores that verify their validity. This approach allowed for the explicit validation of pixels that play a significant role in influencing the classification outcome. Additionally, the quality and features of images markedly affect the performance of algorithms and the precision of their predictions. Hence, the proposed Simple Linear Iterative Clustering (SLIC) augmented with a Gaussian smoothing filter — SLIC-G, served as an image preprocessing technique. It succeeded in segmenting images into regions that convey significant information with a substantially reduced dataset, rather than relying on all the pixels within an image. This method also improved image clarity by removing noise. Experimental outcomes indicated that employing SLIC-G for image preprocessing improved the results of classification.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)