Resumen
Mutated cells may constitute a source of cancer. As an effective approach to quantifying the extent of cancer, cell image segmentation is of particular importance for understanding the mechanism of the disease, observing the degree of cancer cell lesions, and improving the efficiency of treatment and the useful effect of drugs. However, traditional image segmentation models are not ideal solutions for cancer cell image segmentation due to the fact that cancer cells are highly dense and vary in shape and size. To tackle this problem, this paper proposes a novel U-Net-based image segmentation model, named U-Net_dc, which expands twice the original U-Net encoder and decoder and, in addition, uses a skip connection operation between them, for better extraction of the image features. In addition, the feature maps of the last few U-Net layers are upsampled to the same size and then concatenated together for producing the final output, which allows the final feature map to retain many deep-level features. Moreover, dense atrous convolution (DAC) and residual multi-kernel pooling (RMP) modules are introduced between the encoder and decoder, which helps the model obtain receptive fields of different sizes, better extract rich feature expression, detect objects of different sizes, and better obtain context information. According to the results obtained from experiments conducted on the Tsinghua University?s private dataset of endometrial cancer cells and the publicly available Data Science Bowl 2018 (DSB2018) dataset, the proposed U-Net_dc model outperforms all state-of-the-art models included in the performance comparison study, based on all evaluation metrics used.