Research On The Identification Method Of Cervical Cancer Lesions Based On Magnetic Resonance Imaging

Cervical cancer is one of the most common malignant tumors in women worldwide.There are three main ways to prevent cervical cancer: vaccination,screening and treat-ment.Magnetic resonance imaging is considered to be the best imaging technique for cervical diagnosis.Improving the recognition accuracy of magnetic resonance imaging in cervical cancer tumor area is of great significance to reduce the mortality rate of cer-vical cancer in women.Due to the heterogeneity and low contrast of biomedical images,the current cervical cancer lesion identification methods mainly have the following prob-lems:（1）There are few researches on cervical cancer lesion identification,and there is a lack of standard magnetic resonance imaging datasets?（2）The current cervical cancer identification algorithm is not sensitive to the detection of small lesions?（3）The current2 D lesion identification algorithm lacks coherence and consistency in the identification of magnetic resonance images?（4）The 3D lesion recognition network has a large amount of parameters,a large amount of calculation,and the network is difficult to train.This paper studies the above problems,and the specific work is mainly divided into the following points:（1）A standard cervical cancer magnetic resonance imaging dataset was created.The cervical region dataset in this paper contains a total of 894 weighted T2 coronal images,each consisting of 20-35 slices.（2）A set of recognition algorithm Semi Seg Med-Net based on semi-supervised learning is proposed.In this algorithm,a dual-branch structure is designed for pixel-level segmentation and image-level error cor-rection respectively.In the absence of large-scale medical image standard datasets,the semi-supervised recognition method improves the efficiency of image feature utilization,and the performance of the overall indicators is equivalent to that of the fully supervised method.（3）An enhanced 3D multi-scale recognition network Aug MS-Net is proposed.Multi-scale strategy is considered to be one of the effective algorithms to solve the problem of small target recognition.In this paper,an enhanced multi-scale recognition network is proposed to solve the problem of small target recognition.Aug MS-Net reduces the amount of 3D network parameters and improves accuracy and consistency of the recognition re-sults.We evaluate the above methods on two datasets,and the comparison with other state-of-the-art methods shows that our proposed methods achieve steady improvement on var-ious mainstream evaluation metrics while reducing the amount of learnable parameters.Our proposed semi-supervised identification method achieves comparable index results to fully supervised methods on the cervical cancer MRI dataset using a much smaller amount of data.