AI-based Human Cervical Cancer Screening Using Optical Coherence Tomography

As a branch of artificial intelligence,deep learning has achieved great success in computer vision,especially in image classification tasks.Many medical and computer researchers are exploring new areas—using deep learning to help diagnose human diseases.Cervical cancer is one of the most common diseases in women.The existing screening techniques for cervical cancer have different defects,leading to the inability of women to receive treatment in time in many deprived areas.Optical Coherence Tomography(OCT)has been widely used in medical imaging.It can obtain longitudinal images of tissues,provide high-resolution cell-level features,as week as quickly acquire 2D/3D in situ.These advantages make it possible to become a new cervical screening and detection technology.Based on the feature-rich images acquired by OCT system and the powerful capabilities of image analysis of deep learning,this article focuses on proposing a novel cervical disease screening technology to make up for the deficiencies of the existing technology.Based on WinForm C++ and GPU CUDA parallel programs,this paper realized the image acquisition and processing system to control the relevant hardware equipment to scan the tissue to obtain the images.The main tasks accomplished in this paper are as follows:(1)In this thesis,the visual feature extractor of the image is trained by using a convolutional neural network,and the obtained image feature sequence and text feature sequence are combined to train the traditional support vector machine(SVM)model to complete the corresponding classification task.For the collection of 497 3D OCM(Optical Coherence Microscopy)image data sets from 92 Chinese women(each 3D image includes 600 grayscale images)and medical record information(age and HPV test results),we designed 10-fold cross-validation experiments to evaluate the model.On the other hand,we also designed the human blind test.As a result,in the five categories(normal,valgus,low-grade lesions,high-grade lesions,cancer)tasks,the accuracy rate reached 88.3±4.9%(mean ± standard deviation)(average of 74.3%for three human expert tests).In the two-category("low-risk" and "high-risk")missions,the accuracy rate was 91.3 ± 0.051%(average of 88.9%for the three expert tests).(2)In this thesis,the algorithm of deep feature recognition is studied,including the saliency map based on backpropagation,as well as the guided backpropagation a kind of deconvolution.These algorithms are applied to the OCM image classification model of cervical tissue.The experimental results show that these two algorithms can identify typical cervical tissue pathological features,such as normal tissue structure,independent epithelium,squamous cell structure for healthy tissues;and for cancer,cancer cells can be identified.In the future work,based on the experimental results,we can develop cervical cancer screening software that can provide quick diagnosis and feature prompts to assist doctors in diagnosis.