검색 상세

Artificial Intelligence Uncertainty-Based Segmentation for Medical X-ray Images Sang Min Han

  • 한상민 (아주대학교 일반대학원)

초록/요약

Despite technological advances in medical imaging, more sophisticated imaging modal- ities remain expensive and labor-intensive compared to X-rays. Consequently, precise detection and segmentation of pathological regions in X-ray images continue to represent a critical research challenge. Specifically, chest X-ray analysis presents intrinsic complexity arising from the inherent variability in disease morphology and anatomical presentation. Traditional segmentation models encounter substantial limitations in accurately delineat- ing pathological areas, necessitating ongoing methodological refinements. In this study, we propose an Evidential U-Net that leverages evidential losses within a U-Net architecture to enhance the detection and segmentation of diseased areas in chest X-ray images. Training the model with evidential losses enables probabilistic prediction of multi-class distributions, facilitating uncertainty quantification for each output. We integrate evidential losses with focal loss to address class imbalance and dice loss to opti- mize segmentation performance. We evaluate the Evidential U-Net on chest X-ray datasets encompassing patients with COVID-19, tuberculosis, pneumonia, and pneumothorax, sys- tematically comparing its performance against the baseline U-Net model. Keywords: Evidential Neural Networks (ENN), U-Net, X-ray Image Segmentation

more

목차

1 Introduction 1
2 Background 3
2.1 Evidential Deep Learning 3
2.2 U-Net 4
3 Methodology 6
3.1 Datasets 6
3.2 Evidential U-Net Architecture 7
3.3 Loss Function 8
3.3.1 Evidential loss 9
3.3.2 Dice loss 9
3.3.3 Focal Loss 10
3.3.4 Overall Loss 11
3.4 Evaluation Metrics 11
4 Experiments and Results 13
4.1 Experimental Settings 13
4.2 Performance comparison 14
4.3 Empirical Studies 15
5 Discussion 19
6 Conclusion 21

more
반출 Meta View 목록