| dc.description.abstract | Medical image analysis has advanced rapidly with the integration of deep learning techniques. However, the challenge of unbalanced datasets and the need for effective pre-processing methods remain significant difficulties in achieving optimal classification performance. This thesis aims to investigate the effectiveness of various image dataset augmentation techniques and the potential of diffusion models for chest x-ray image classification, focusing mainly on the data unbalance problem.
After obtaining a high quality dataset using various data and image preprocessing methods,
we used traditional data augmentation methods such as rotation, flipping, blurring, and
contrast modification to increase the number of positive class samples. In addition to
traditional augmentation methods, several diffusion models are introduced to synthesize
new chest x-ray images to further strengthen the minority class and address data imbalance. The performance of these methods was evaluated using specific metrics and compared to established baseline models.
The results showed that while replacing labels and masking images can introduce errors, the selected combination of preprocessing methods showed promise in improving classification performance. These results also indicated that traditional data augmentation methods, after careful fine-tuning of the hyperparameters, achieved significant performance improvements over the original baseline model. The application of diffusion models further improves the final classification results. Moreover, the images generated by the diffusion models, compared to traditional augmentation methods, do not merely modify the original images, but introduce some new image information, leading to improvements in various metrics.
Our results demonstrate the importance of augmentation methods in addressing data imbalance and improving the results of chest X-ray image classification. The research describes the most important image generation techniques that yield superior classification results while overcoming the hurdles of imbalanced datasets. These findings have profound implications for the medical field and machine learning specialists, signaling a promising path for improving diagnostic accuracy and patient care in chest X-ray image analysis. While current methods show potential, further research, particularly in the areas of stable diffusion models and deep learning-based image classification, is needed to make significant
advances in the field. | |
