Research On Cervical OCT Image Classification And Lesion Detection Based On Dual Attention Mechanism

Optical coherence tomography(OCT)has high-efficiency and non-invasive characteristics and has been used in the screening and diagnosis of ophthalmology,cardiovascular and other diseases.In recent years,OCT technology has also been gradually applied to examining and diagnosing cervical diseases.However,for gynecological clinicians,the application of new technologies has increased the cost of learning.In the classification and segmentation tasks of fundus OCT images,the existing auxiliary diagnosis methods based on deep learning have been clinically recognized.Therefore,applying deep learning technology to complete the classification of cervical tissue OCT images and the detection of lesion areas has become a topic worthy of research.There are the following main problems in data and algorithms to develop a computer-aided system based on cervical tissue OCT images.1)Due to the lack of standardized and labeled datasets,labeling data still depends on professional physicians,but their time is limited.It is an urgent need to improve the labeling efficiency of professional physicians.2)Designing a deep neural network model and training strategy for labeled OCT images to obtain the results recognized by professional doctors is also an urgent problem to be solved.To solve the above problems,we introduce the idea of “human-in-the-loop” in the labeling and training process.We train our deep neural network model through the iteration between model prediction and doctors’ error correction.By continuously accumulating labeled data,the model and data can be improved together.Consider that convolutional neural networks have achieved great success in computer vision tasks in recent years.This paper introduces two attention mechanisms:channel attention and spatial attention,into the convolutional neural network to better mine the correlation between image features.We adopt the image classification method based on image patches to apply deep learning technology to real clinical scenes.We split the entire OCT image into multiple patches,which facilitates the detection of the focal points and reduces the distortion of the tissue morphology caused by image processing to a certain extent.On this basis,we propose and test a variety of voting mechanisms for label prediction of the entire image to give the final auxiliary diagnosis result.Considering the interpretability of assistant diagnosis results,we also combine some deep-learning interpretability methods to visually display the model’s focus,assisting professional physicians in making correct judgments.In this paper,we carry out our experiments on 862 cervical tissues OCT images of450 patients provided by a partner hospital.The designed model can achieve 95.8%classification accuracy on high-risk prediction tasks through the ten-fold crossover verification.With a sensitivity of 100.0 %,the specificity reaches 94.0%,which meets clinical application requirements.Simultaneously,visual aids have also been developed to help professional physicians make diagnoses more conveniently.