A lightweight PCT-Net for segmenting neural fibers in low-quality CCM images.
Journal:
Computers in biology and medicine
PMID:
40121803
Abstract
In this paper, we propose a lightweight Position Channel Transformer Network (PCT-Net) for segmenting slender neural fibers in low-quality corneal confocal microscopy images with speckle noise and uneven lighting. Three modules including the channel attention (CA) module, the dilated shuffle positional (DSP) module, and the shuffle contextual transformer positional (SCTP) module are proposed and stacked to form the PCT-Net. The proposed channel attention module captures global information and adaptively learns to identify and emphasize key channel features in corneal confocal microscopy images by utilizing global average pooling and global maximum pooling. The dilated shuffle positional module enhances the model's ability to recognize local details using the dilated convolution and the positional focus mechanism. The shuffle contextual transformer positional module further integrates the contextual transformer (CoT) module, which enables the network to understand the image content more comprehensively by fusing global and local contextual information. In addition, PCT-Net integrates features from different stages by adding skip connections to prevent the loss of detailed fiber information. In particular, it adopts a dual output training strategy, effectively integrating multi-scale features. We then address the imbalance between labeling categories in the dataset by using morphological expansion operations on structural elements to extend vessel labels. In addition, we propose a hybrid loss function that combines multi-scale structural similarity index measure (MS-SSIM) loss, binary cross entropy (BCE) loss, and polyfocal loss to train PCT-Net. With the proposed dataset and loss function, we train the PCT-Net successfully. We then test our method on three publicly available datasets and compared it with seven existing image segmentation methods. The experimental results show PCT-Net's superior performance in segmenting low-quality corneal confocal microscopy images with speckle noise and uneven illumination. It gets better results than the compared methods. Furthermore, the proposed PCT-Net has fewer parameters, which provides the possibility for portable devices to perform real-time image segmentation.