Deep learning-based super resolution for enhancing of resolution in low magnification of microscopic sperm images in pigs

초록

Background: The morphological characteristics of sperm are essential for assessing sperm quality, which partially influences reproductive efficiency variables such as litter size in sows. Recently, deep-learning-based object detection algorithms have been explored to detect and classify sperm morphological features, with the training of these models requiring sperm microscopy image data. The performance of these models in detecting morphological features was significantly affected by the size of the dataset and the image quality of the images. Methods: This study proposed a deep-learning-based super-resolution (SR) algorithm to enhance the quality of sperm microscope images. The model was trained using a dataset consisting of high-resolution (HR) original and low-resolution (LR) images generated by downscaling the original images through bicubic interpolation. The SR results of the test dataset were compared with those of the original HR images to evaluate their performances using the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Statistical analyses were conducted to examine the performance differences based on the training models and dataset types. Results: This study indicated that the SR images showed no significant differences in PSNR (p = 0.9740) and SSIM (p = 0.9864) compared with the original HR images in same magnification. Moreover, increased processing speed was observed with reductions in model hyperparameters. While processing SR images improved spatial resolution across various microscopic magnifications, the overall image quality did not exceed that of the original HR images. Conclusions: SR models applied to sperm microscopy images outperformed conventional SR algorithms. These findings suggest that SR algorithms hold promise for improving the quality of LR microscopic images in future deep-learning-based object-detection algorithms.

키워드