Toward Advanced Deep Learning Techniques for Medical Image Analysis

Abstract

Deep learning (DL) has revolutionized medical diagnosis by enabling the analysis of diverse data modalities, particularly medical images, from a wide range of sources. This transformative capability has made DL approaches an increasingly vital tool in modern healthcare. Indeed, the potential of DL in medical image analysis is undeniably promising, with significant advancements already being made. However, only a few DL-based approaches have successfully transitioned into clinical practice. This may be due to various factors such as overfitted models, selection bias, and the extensive preprocessing of datasets, which fail to accurately represent clinical diversity and local variations. This thesis aims to explore the transformative potential of DL in medical image analysis, focusing on the development of novel models that enhance diagnostic accuracy, efficiency, and personalization. In this thesis, we address several key medical imaging modalities—X-rays, computed tomography (CT) scans, optical coherence tomography (OCT), and dermatoscopic imaging—to tackle critical diagnostic challenges in diseases such as COVID-19, retinal disorders, and skin lesions. This work offers a thorough exploration of advanced DL architectures and hybrid methodologies tailored for medical diagnostics. The first contribution underscores the effectiveness of attention mechanisms in achieving high-accuracy diagnoses for COVID-19 and other pulmonary diseases. Specifically, the integrated attention mechanisms resulted in notable improvements, with quantified performance metrics indicating high sensitivity and precision as high as 99% in COVID-19 diagnosis. The second introduces HTC-Retina, a cutting-edge hybrid approach that combines vision transformers (ViTs) with convolutional neural networks (CNNs). This approach, applied to OCT image classification, showed enhanced performance in identifying retinal disorders, with significant improvements in accuracy over traditional methods, achieving classification rates ranging from 97% to 99%. The third presents JILDYA-Net, a novel lightweight model optimized for skin lesion classification. This model employs an efficient architecture designed for resource-constrained clinical settings. It was evaluated on dermatoscopic images and demonstrated superior classification accuracy. The final contribution is the development of DA-UNet-Plus, a novel hybrid segmentation model. Initially applied to detect intraretinal fluid in OCT images with notable performance, DA-UNet-Plus was later adapted for skin lesion localization in dermatoscopic images, achieving improved lesion boundary delineation. Overall, these contributions demonstrate the powerful potential of DL in automating complex diagnostic tasks and fostering personalized healthcare while paving the way for future innovations in medical image analysis.