Doktorandské kolokvium KAI - Thakar Mayur Pramod (25.5.2026)
v pondelok 25.5.2026 o 13:10 hod. v miestnosti I 9
Prednášajúci: Thakar Mayur Pramod
Názov: Literature Review and Research Directions in Trustworthy AI for Medical Imaging
Termín: 25.5.2026, 13:10 hod., I/9
Abstrakt:
Artificial intelligence has become increasingly important in medical imaging applications such as image segmentation, classification, localization, and computer-assisted diagnosis. Although recent deep learning approaches have achieved high performance across multiple medical imaging tasks, several challenges still limit their reliability and broader clinical adoption. Current research highlights important issues related to explainability, robustness, annotation quality, model generalization, and trustworthiness in AI-assisted medical systems.
This presentation summarizes the initial literature review conducted in the area of trustworthy AI for medical imaging. The review focused on understanding common deep learning workflows used in medical image analysis, including preprocessing, region of interest localization, segmentation, classification, and evaluation methodologies. Particular attention was given to explainable AI approaches, segmentation architectures such as U-Net and nnU-Net, transformer-based models, and hybrid deep learning frameworks applied in CT and CBCT imaging studies.
The reviewed literature indicates that while many existing models report high quantitative performance using metrics such as Dice score, IoU, and classification accuracy, several open research problems remain insufficiently addressed. These include limited dataset availability, annotation inconsistency, lack of robustness across different imaging conditions, and limited validation of explainability methods for real clinical use.
Based on these observations, the presentation outlines possible future research directions related to trustworthy and explainable AI systems in medical imaging, with a particular interest in model reliability, preprocessing workflows, segmentation robustness, and interpretable deep learning methods for clinical applications.