Data Augmentation of Sperm Images Using Generative Adversarial Networks (WGAN-GP)

2025-09-09T20:47:43+00:00

This study analyzes the use of WGAN-GP for data augmentation in the analysis of sperm morphology. WGAN-GP has been the focus in this study for generating sperm microscopy images, which in turn aims to mitigate the problem of data scarcity in medical imaging. A heterogeneous dataset with mixed object categories was initially employed, leading to an FID score of 134, which in turn reflected a high incidence of mode collapse. For this reason, the dataset was divided into subcategories of Normal, Abnormal, and Non-Sperm identifications, with the scores of the subcategories being 59.19, 74.92, and 83.56, respectively, and showing better balanced model stability. This study's primary contribution is the use of WGAN-GP for the first time for sperm image data augmentation and the generation of more realistic synthetic images. Furthermore, this study illustrates the first understanding of the intricacies of data distribution's complexity and its effect on the model's performance, indicating the possibility of improvement using class-based techniques and sophisticated architectures for the generator. The innovation of this study is the application of WGAN-GP to sperm morphology datasets, improving image quality and the stability of the results, coupled with extensive model performance analysis and providing a further understanding of the field of medical image data augmentation.

Unsupervised Optimization of Boundary Information Based on the Coefficient of Variation to Improve Image Segmentation

2025-11-29T14:59:11+00:00

The automatic retrieval of boundary information from image objects suffers from the problem of under and over-segmentation, where the former leads to missed object detection, while the latter delivers an improper object shape. A method to optimize the automatic retrieval of complete and proper boundary information is proposed in this research based on an unsupervised approach. The strategy is to utilize the trade-off between the coefficient of variation from shape distribution against the mean of entropy contribution from segmented regions. This mechanism relies on the assumption that the segmentation result of a natural image contains a prominent main object representation with its details which are presented as a normal distribution of segmented regions. The research also enhances the entropy-based segmentation evaluation by redefining the computation of image entropy and segmentation entropy. The experiment shows that the proposed approach is capable of reducing over-segmentation by 57.20% compared to the existing algorithm, while at the same time reducing the consumption time by 85.26%. The empirical evaluation shows that the proposed approach delivers the highest accuracy among other evaluated methods. Qualitative validation based on groups of human observers shows that the proposed approach is the most desired algorithm for producing boundary information and measuring segmentation quality. These findings suggest that the trade-off between the mean of entropy contribution from the segmented regions and the coefficient of variation from shape distribution becomes an effective feature for unsupervised retrieval of boundary information.

Register: Jurnal Ilmiah Teknologi Sistem Informasi

Data Augmentation of Sperm Images Using Generative Adversarial Networks (WGAN-GP)

Unsupervised Optimization of Boundary Information Based on the Coefficient of Variation to Improve Image Segmentation