Development of Computer-Aided Diagnosis Algorithm(S) for Medical Image Analysis

Abstract

Healthcare serves as a fundamental pillar of human well-being, encompassing the prevention, diagnosis, and treatment of diseases. However, various challenges, including excessive costs, limited resources, and inadequate infrastructure, hinder its accessibility and overall efficiency. The integration of artificial intelligence (AI) into medical sciences has the potential to transform healthcare by enhancing precision, efficiency, and personalization. This thesis examines AI-driven diagnostic systems to address critical gaps in the early detection of breast cancers, colorectal cancer and segmentation of burn tissues, introducing innovative methodologies to automate and improve diagnostic workflows. This thesis is structured into five chapters. The introductory chapter establishes the theoretical foundations, identifies existing research gaps, and defines the overarching objective: the development of robust diagnostic frameworks capable of accurately and effectively categorizing medical images as either healthy or diseased one. Second chapter is related to the diagnosis of breast cancer as breast cancer incidence in India as well as world is rising, necessitating early and accurate detection to improve survival rates. This study proposes a novel ensemble classification framework for breast cancer detection using optical coherence tomography images. Given the variability in datasets and performance metrics, a single classifier may be insufficient. Thus, we integrate the technique for order of preference by similarity to ideal solution for multi-criteria decision making with the crow search algorithm to optimize classifier selection and weight assignment. Additionally, SHapley Additive exPlanations values are employed to enhance model interpretability by visualizing feature attributions. This framework aims to reduce reliance on skilled pathologists, minimize interobserver variability, and accelerate breast tissue assessments. Third chapter belongs to accurate assessment of burn tissue. Accurate assessment of burn injuries is critical for effective treatment and can significantly impact patient outcomes. However, the complexity of burn injuries makes timely and precise diagnosis challenging, particularly in differentiating burn depth. In severe cases, segmentation of burn tissue is essential, yet manual segmentation of three-dimensional (3D) optical coherence tomography datasets is highly time-consuming. Traditional deep learning methods often process each scan in isolation, overlooking valuable inter-slice or longitudinal information that could enhance segmentation accuracy. To address this limitation, we propose a Bi-directional Long Short-Term Memory UNET (Bi-LSTM-UNET) for OCT segmentation, leveraging inter-slice dependencies while maintaining computational efficiency. Experimental results demonstrate the model’s effectiveness, achieving high recall, precision, accuracy, Dice Coefficient, and Intersection over Union (IoU). The developed model has significant learning potential and can assist surgeons by providing a rapid second opinion on burn tissue assessment. Furthermore, its application reduces the workload on medical professionals and can be extended to other medical imaging domains. Fourth chapter is focused on diagnosis of colorectal cancer. Colorectal cancer (CRC) is a leading cause of cancer-related mortality worldwide, and its early and accurate diagnosis is critical for improving patient outcomes. Traditional diagnostic methods, such as histopathological analysis, rely on manual examination, which is time-consuming, subjective, and prone to interobserver variability. Artificial intelligence (AI) has emerged as a powerful tool to enhance diagnostic accuracy, efficiency, and consistency in CRC assessment. This research presents an advanced AI-driven histopathological image classification system that integrates spectral and spatial features to improve colorectal cancer diagnosis. Convolutional neural networks (CNNs) extract spatial features, while a gray-level co-occurrence matrix (GLCM) captures texture information. A multi-resolution wavelet transform is employed for spectral feature extraction. Classification is performed using a random forest for spectral-spatial features and a support vector machine (SVM) for spatial features, with classifier weights optimized through the equilibrium optimization technique. The concluding chapter provides a comprehensive summary of the key findings presented in this thesis and outlines potential future directions for advancing artificial intelligence (AI) in medical imaging. While AI has significantly contributed to improving diagnostic accuracy, efficient segmentation and automation in healthcare, several critical challenges still remains. These include the need for large, diverse, and high-quality datasets to enhance model generalizability, as well as the necessity of improving model interpretability and explainability to foster trust among medical professionals and regulatory bodies. Future research will focus on addressing these limitations by developing more robust and ethically responsible AI models, incorporating advanced techniques for explainable AI (XAI), and ensuring equitable AI deployment across diverse populations. Additionally, interdisciplinary collaboration between AI researchers, clinicians, and policymakers will be essential to facilitate the seamless integration of AI-driven diagnostic systems into clinical practice. By overcoming these challenges, AI can be more effectively leveraged to revolutionize medical imaging, ultimately leading to improved patient outcomes and more accessible healthcare solutions.