Biomedical image segmentation with modified U-Net

Image segmentation has a very significant place in image processing and machine vision. Exclusively, it plays a vital role to improvement of models to help specialist in biomedical and medical areas, saving time and cost in existing methods. Among the most significant image segmentation models utilize deep learning at U-Net model. U-Net model has motivated numerous studies and has been a useful model in development of convolutional neural networks. Scope of the report, we suggest a new version of U-Net, U-Net 11, which utilizes 11 convolutional layers and assert a few modifies to ameliorate to segmentation accuracy. The standard U-Net method was ameliorated and checked on three types of datasets and outperformed the standard U-Net model. U-Net 11 model: Breast tumor cell segmentation from cell tissues, lung tumor segmentation from CT scan figures and cell nucleus segmentation at the Data Science Bowl 2018 rivalry. Abovementioned datasets are significant owing to their alteration quantity of images and changing levels of challenges in segmentation tasks. On the purpose of observe the layer variation, U-Net 13 and U-Net 7 models are generated by adding and removing layers to standard U-Net and compared with the suggested method. A modified U-Net model: 69.09%, 95.02%, and 81.10% DSC accuracy on the breast cancer cell segmentation dataset derived from cell tissues, lung segmentation dataset obtained from computed tomography images, and cell nucleus segmentation dataset obtained from tissue images, which are respectively 5%, 2%, and 4% higher besides standard U-Net model. This variation in accuracy rates is especially important for critical segmentation datasets in biomedical engineering and medicine, and model can be improved to achieve more successful results, leading to significant savings in these fields. Keywords: U-Net, image segmentation, deep learning, biomedical image