Year: 2025
Type: article
Abstract: In this study, we investigate the effectiveness of different deep learning architectures in the task of medical image synthesis using convolutional neural networks. Our goal is to compare the performance of standard ResNet architectures (ResNet-18 and ResNet-50) with their Squeeze-and-Excitation (SE) enhanced counterparts (SE-ResNet-18 and SE-ResNet-50). The evaluation is conducted on three publicly available medical datasets: CVC-ClinicDB (colorectal polyp images), Messidor2 (retinal images), and Pap Smear (cervical cell images). For image synthesis, we employ these architectures as generative backbones and assess the quality of the generated images using both pixel-level metrics Mean Squared Error (MSE) and perceptual similarity metrics, namely Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). Experimental results demonstrate that SE-enhanced ResNet architectures outperform their vanilla counterparts in generating more realistic and perceptually coherent images. Particularly, SE-ResNet-50 achieves the lowest FID and KID scores across all datasets, indicating superior generative quality. These findings highlight the impact of channel-wise attention mechanisms in enhancing feature representation and improving medical image synthesis tasks. Experimental results demonstrate that ResNet50 achieves the best performance across multiple metrics, including LPIPS, FID, KID, and MSE, confirming its superiority in both perceptual quality and pixel-level accuracy.
Institution: Tarsus University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7217192868