Biyomedikal görüntülerin segmentasyonu için derin öğrenme yaklaşımları

Motivated by the crucial role of biomedical image segmentation in diagnostic workflows, this thesis investigates the potential of deep learning-based models across three distinct and challenging clinical domains. The study was conducted across three clinical scenarios; segmentation of microbial dental plaque from intraoral camera images, low grade glioma tumors from MR, and the prostate gland from PET/CT. For each task, deep learning models were rigorously compared against classic computer vision methods, with systematic hyperparameter optimization performed to ensure a fair evaluation ground. The findings consistently demonstrated the overwhelming superiority of deep learning models over classical methods across all three domains. Notably, the U-Net Transformer model achieved statistically significantly better results than three expert dentists in dental plaque segmentation, underscoring this technology's clinical potential. Similarly, the UNet++ and Attention U-Net architectures delivered the highest performance for low-grade glioma and prostate segmentation, respectively. Furthermore, explainable AI techniques, specifically Grad-CAM and Saliency Maps, confirmed that the decision mechanisms of these high-scoring models correctly focused on the targeted anatomical regions. This process has successfully moved the models beyond the black box perception. This thesis provides strong evidence that task-specific deep learning solutions across diverse medical imaging modalities can not only enhance diagnostic accuracy beyond expert-level performance but also improve clinical efficiency through automation and explainability.