Akademik Veri Yönetim Sistemi
2026 International Conference on Advances in Artificial Intelligence and Machine Learning, AAIML 2026, Tokyo, Japonya, 20 - 22 Mart 2026, ss.447-452, (Tam Metin Bildiri)
In this paper, we present URes2Net, a novel and efficient architecture developed for medical image segmentation. The architecture of our model is based on a U-shaped network, in which both the encoder and the decoder part are improved by a proposed residual block, called Res2Block. Each Res2Block consists of a one-stage encoder-decoder pair to capture multiscale contextual information. Furthermore, the backbone of each Res2Block is built upon Res2Net, which enables hierarchical residual connections within the block. This design allows the network to effectively capture local and global contextual information while extracting features at different scales. URes2Net was evaluated on the ISIC-2017 skin lesion segmentation dataset. Quantitative results demonstrate that our model achieves a Dice loss of 0.819, MAE of 0.084, and a maximum F-measure of 0.897, which clearly outperforms comparable U-shaped architectures. Qualitative results show that the predicted probability maps align closely with ground truth masks across lesions of varying sizes. Thus, the network preserves fine details while accurately segmenting complex structures. By integrating multi-scale contextual information through hierarchical residual connections, the network demonstrates good results in detecting lesions with diverse shapes and sizes. The proposed model can also be extended to other biomedical image segmentation tasks, making it a promising solution for real-world clinical applications. The source code is available: https://github.com/sym-codes/URes2Net.