سالار زارعي

In medical image processing, access to high-quality an‎d diverse datasets remains a significant challenge, as collecting large-scale data is often difficult, time-consuming, an‎d costly. To address this limitation, the U-Net architecture was proposed, which enables accurate results even when trained on relatively limited datasets. The development of deep convolutional neural networks for image segmentation has consistently faced the challenge of balancing model depth with computational efficiency. A key concern in this area is achieving deeper network layers without disproportionately increasing computational costs while maintaining the resolution of feature maps. To address this issue, the U2-Net architecture was introduced, employing a nested U-shaped, two-level structure that allows training a deep network from scratch without relying on pre-trained weights from image classification tasks. In this study, sharpening blocks were incorporated into U2-Net, applying a depthwise convolution with a sharpening kernel to encoder feature maps before merging them with decoder features. This process produces an intermediate sharpened feature map of the same size as the encoder map, enabling the integration of semantically diverse features an‎d reducing noise in the early network layers, without introducing additional trainable parameters. In this study, by appropriately combining the U2-Net an‎d Sharp U2-Net models, experiments were conducted on the Lung Segmentation an‎d ISIC-2018 datasets. The eva‎luations demonstrate that the proposed models outperform recent state-of-the-art methods in medical image segmentation tasks. Notably, Sharp U2-Net+ achieves a 0.91% improvement in Dice coefficient an‎d a 1.84% increase in IoU on the Lung Segmentation dataset, as well as a 3.44% improvement in Dice coefficient an‎d a 2.82% increase in IoU on the ISIC-2018 dataset, while reducing the approximate memory footprint by 102 MB compared to one of the best recent models.

فيلتر تيزكننده , بخش‌بندي معنايي , شبكه كاملاً كانولوشني , اتصالات گذري

U-Net

Salar Zarei

Dr. Ali Sadr