Development of New Framework for Medical Image Analysis

Abstract

Healthcare is a cornerstone of human welfare, encompassing the prevention, diagnosis, and treatment of diseases. However, challenges such as exorbitant expenses, limited resources, and insufficient infrastructure hinders its accessibility and efficiency. Integrating artificial intelligence (AI) into medical sciences has the potential to revolutionize healthcare by improving precision, efficiency, and personalization. This thesis explores AI-driven diagnostic systems to address critical gaps in the early detection of cancers and blood disorders, presenting innovative approaches to automate and enhance diagnostic workflows. The thesis consists of six chapters. The introductory chapter establishes the theoretical foundations, highlights research gaps, and defines the overarching goal: developing robust diagnostic frameworks capable of effectively and precisely categorizing medical images as healthy or diseased. Attention is then paid to haematological malignancy, particularly diffuse large B cell lymphoma (DLBCL), a type of blood cancer that requires timely detection and intervention in mitigating the morbidity and mortality. To overcome the limitations of traditional diagnostic techniques, an AI- powered deep discriminative learning model (DDLM) with calibrated attention maps (CAM) is proposed. This system employs histopathological image analysis to differentiate DLBCL from non-DLBCL, effectively addressing inter-class variability and intra-class similarities. A comparative evaluation demonstrates the model’s efficacy, with its utility and adaptability in diverse clinical scenarios also highlighted. Next focus is given on breast cancer (BC)- a prevalent malignancy in female population ranking 2nd in terms of lethality. Conventional treatment protocols often involve partial or complete removal of breast tissue along with surgical excision of the tumor. However, these interventions frequently fail to achieve complete eradication of the tumor. Histological assessment of breast tissue, while widely recognized as the gold standard for diagnosis, is labor-intensive, time- consuming, and impractical for real-time use during surgery. To address these limitations, this study introduces a cutting-edge imaging system based on full-field polarization-sensitive optical coherence tomography (FF-PS-OCT) integrated with an ensemble learning model, optimized using the technique for order preference by similarity to ideal solution (TOPSIS) ranking method. This approach offers a rapid and accurate ex-vivo alternative to traditional histology, enhancing intraoperative decision-making and reducing recurrence rates. The next chapter presents a high-resolution intelligent system designed for diagnosing sickle cell disease (SCD), a hereditary blood disorder that is highly prevalent in the Caribbean as well as in Western and Central sub-Saharan Africa. Ranked as the 12th leading cause of death globally, SCD presents significant diagnostic challenges, particularly in resource-limited settings. This study proposes an automated solution using a 3D intelligent quantitative phase microscope to detect sickle cells in blood smears. The system enhances diagnostic efficiency by significantly improving speed and accuracy while reducing reliance on expert resources. The ensuing chapter deals with the enhanced and explainable renal histopathology image classification using a model that incorporates advanced techniques. Renal cancer- ranked as the 9th most common ailment, poses significant diagnostic challenges due to its diverse subtypes and asymptomatic progression. This study proposes a novel end-to-end learning approach for classifying clear cell renal cell carcinoma using a deep discriminative model which incorporates a global- local attention network. The integration of SHapley Additive exPlanations for explainable AI plays a crucial role in enhancing the interpretability of model predictions. The model effectively tackles class ambiguities and achieves high accuracy in histopathological image classification.

The concluding chapter encapsulates the key findings of this thesis and highlights future directions for advancing AI in medical imaging. While AI has undeniably transformed healthcare by enhancing diagnostic accuracy and efficiency, significant challenges persist, such as the necessity for large, diverse datasets and improved model explainability. Future efforts will focus on addressing these obstacles to ensure the broader adoption of AI across varied populations and building trust through greater transparency and explainability. This thesis underscores the transformative potential of AI in revolutionizing healthcare, offering innovative solutions to critical challenges in disease detection. By advancing precision, efficiency, and accessibility, the work paves the way for a future where AI-driven diagnostics are seamlessly integrated into global healthcare systems, reshaping clinical practices and improving patient outcomes.