Dual-attention EfficientNet based on multi-view feature fusion for cervical squamous intraepithelial lesions diagnosis

Cervicograms are widely used in cervical cancer screening but exhibit a high misdiagnosis rate. Even senior experts show only 48% specificity on clinical examinations. Most existing methods only use single-view images applied with acetic acid or Lugol's iodine solution as their input data, ignoring the fact that non-pathological tissues may show false-positive reactions in these single-view images. This can lead to misdiagnosis in clinical diagnosis. Therefore, it is essential to extract features from multi-view colposcopy images (including the original images) as inputs, because three-view cervicograms provide complementary information. In this work, we propose an improved EfficientNet based on multi-view feature fusion for the automatic diagnosis of cervical squamous intraepithelial lesions. Specifically, EfficientNet-B0 is employed as the backbone network, and three-view images are taken as inputs by channel cascading to reduce misclassification. Additionally, we propose a dual-attention mechanism that implements the feature selection function based on Convolution Block Attention Module (CBAM) and Coordinate Attention (CA). These two attention mechanisms assist each other to enhance the feature representation of HSIL. We leverage a dataset of 3294 clinical cervigrams and obtain 90.0% accuracy with recall, specificity, and F1-Score of 87.1%, 93.0%, and 89.7%, respectively. Experimental results prove that this method can help clinicians with precise disease classification and diagnosis, and outperforms known related works.