A novel image segmentation network with multi-scale and flow-guided attention for early screening of vaginal intraepithelial neoplasia (VAIN).

Background: Vaginal intraepithelial neoplasia (VAIN) is a rare precancerous lesion, and early diagnosis is crucial for preventing its progression to invasive vaginal cancer. However, the subtle differences in morphology and color between VAIN lesions and normal vaginal tissue make the automatic segmentation of VAIN highly challenging. Existing methods struggle to achieve precise segmentation, impacting the efficiency of early screening.

Purpose: This study aims to develop a high-accuracy, robust deep learning image segmentation network to accurately and automatically segment VAIN lesions, thereby improving the efficiency and accuracy of early VAIN screening.

Methods: We propose a multi-scale dilated attention flow network for VAIN image segmentation. This network improves upon the U-Net architecture by optimizing the designs of the encoder and decoder and incorporating skip connection modules. In the encoding stage, we introduce the dilated squeeze-and-excitation (DiSE) module and the flow field guided adaptive separation and enhancement (FGASE) module. The DiSE module integrates dilated convolutions with varying dilation rates and a channel attention mechanism, effectively extracting multi-scale contextual information and enhancing the model's ability to perceive VAIN lesions of different sizes. The FGASE module employs flow-guided techniques to dynamically separate the features of the main region (VAIN lesions) from the edge region and enhance them individually. In the decoding stage, we propose a depth wise enhanced pooling (DEP) module that combines deep convolutional layers with adaptive pooling strategies to improve local feature extraction capabilities and optimize global contextual information. The skip connection stage introduces a triple statistical attention (TSA) module that utilizes global average pooling, global max pooling, and global standard deviation pooling to effectively capture diverse feature information, thereby enhancing the model's ability to model long-range dependencies.

Results: Experiments conducted on a VAIN image dataset comprising 1142 patients demonstrate that the proposed network significantly outperforms other medical image segmentation methods across six metrics: Mean intersection over union (MIoU), dice coefficient, accuracy, recall, precision, and mean absolute error (MAE). Specifically, this network achieved an MIoU of 0.8461 and a Dice coefficient of 0.9166, substantially higher than other comparative methods, with a faster convergence speed. Ablation studies further confirm the effectiveness of each module in enhancing the model's performance.

Conclusions: The proposed network exhibits exceptional performance and robustness in the task of VAIN image segmentation, effectively segmenting VAIN lesions and providing strong technical support for early VAIN screening and clinical diagnosis. This work has significant clinical application value.