<![CDATA[Medical image analysis for semantic segmentation using deep learning technology has been extensively developed. One of the notable architectures is U-NET, which has demonstrated high accuracy in segmentation tasks. Further advancements have led to the development of ELU-NET, which aims to enhance model efficiency. ELU-NET achieves relatively good accuracy; however, further comparative analysis of both models is necessary. The comparison between these models is based on accuracy, storage usage, and processing time in performing semantic segmentation of pancreatic cancer images. The pancreatic cancer images utilized in this study are sourced from the PAIP 2023 Challenge, consisting of hematoxylin and eosin (H&E)-stained images. Experiments were conducted by varying the number of filters and model depth for both architectures. The evaluation was performed using a dataset of 57 pancreatic cancer images. The experimental results indicated that U-NET achieved the highest accuracy at 92.8%, slightly outperforming ELU-NET, which attained 89.7%. However, ELU-NET is significantly more efficient in terms of storage usage (8.1 MB for ELU-NET compared to 93.31 MB for U-NET) and processing time (4.0 s for ELU-NET and 5.3 s for U-NET). Although ELU-NET exhibited slightly lower accuracy than U-NET, it surpassed U-NET considerably in terms of storage efficiency (by 85.21 MB) and processing speed (by 1.3 s). These findings suggest that ELU-NET is not superior to U-NET in accuracy. However, given the storage size ratio of 1:11.51 and the processing time ratio of 1:1.325 between ELU-NET and U-NET, the 3.1% accuracy difference represents a reasonable trade-off.]]>
Copyrights © 2025
