PGC-Net: A Novel Encoder-Decoder Network With Path Gradient Flow Control for Cell Counting

To carry out cell counting, it is common to use neural network models with an encoder-decoder structure to generate regression density maps. In the encoder-decoder structure, skip connections are usually used to retain detailed features. However, skip connections introduce to the encoder multiple reverse propagation paths; the backward propagation gradients along these paths exhibit significant semantic differences, which affect the encoder’s training process and may lead to adverse effects. To remedy this problem, we propose a path-gradient controlling network for cell counting. First, a novel reverse gradient control module is proposed to balance the impact on the encoder of the backward propagation signal from the skip connections. Second, to eliminate noise in the feature maps of the encoder output, the convolutional and channel attention modules are used on the shallowest layer’s skip connection. Finally, we utilise depthwise convolution to reduce information loss during the downsampling process, and we use depthwise separable transposed convolution as the upsampling method to mitigate overfitting. Experiments demonstrate that the proposed method outperforms state-of-the-art techniques such as MSCA-UNet, Two-Path Net, SAU-Net, and Cell-Net in terms of the mean absolute error (MAE) metric on four publicly available cell-counting benchmark datasets. Our model performs better on the synthetic bacterial (VGG) dataset (1.9 ± 0.1) than does the MSCA-UNet (2.0 ± 0.2). On the Modified Bone Marrow (MBM) dataset, our model (3.7 ± 0.2) outperforms SAU-Net (5.7 ± 1.2). On the human subcutaneous adipose tissue (ADI) dataset, our model, with (8.9 ± 0.3), surpasses MSCA-UNet with (9.8 ± 0.7). On the Dublin Cell Counting (DCC) dataset, our model achieves (2.4 ± 0.2) and outperforms SAU-Net with (3.0 ± 0.3). The source code of our method is available at
https://github.com/mona-aliye/PGC-Net
.