1. Sharp dense U-Net: an enhanced dense U-Net architecture for nucleus segmentation.
- Author
-
Senapati, Pradip, Basu, Anusua, Deb, Mainak, and Dhal, Krishna Gopal
- Abstract
Deep Learning-based algorithms have shown that they are the best at segmenting, processing, detecting, and classifying medical images. U-Net is a famous Deep Learning (DL) approach for these applications. U-Net conducts four down-samplings before the concatenate process, resulting in low resolution. The dense U-Net design overcomes this problem, but the huge semantic gap between low-level and high-level down-sampling and up-sampling features remains a key concern. This work designed Sharp Dense U-Net, an improved U-Net architecture for nucleus segmentation, to solve these constraints. In the down-sampling path, dense and transition operations are used instead of max pooling and convolution to extract more informative information. In the up-sampling path, a new up-sampling layer, merging, and dense blocks reconstitute high-resolution images. Sharpening spatial filters take the place of skip connections to stop feature mismatches between the decoder and encoder paths. The proposed model is trained on the combined dataset and obtains dice coefficients, IoU, and accuracy of 0.6856, 0.5248, and 84.49, respectively. For nucleus segmentation from histopathology images, the Sharp Dense U-Net model is better than the U-Net, Dense U-Net, SCPP-Net, and LiverNet. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF