Research For Segmentation Of Cervical Cancer Tumor Based On Deep Learning

Cervical cancer is one of the most common malignant tumors worldwide.Accurate detection of tumor from magnetic resonance image is critical for clinical diagnosis and treatment.With the rapid development of deep learning technology,many methods based on deep neural networks have achieved good effects in medical image segmentation tasks.Due to the heterogeneity of medical images,the current cervical cancer tumor segmentation methods mainly have the following problems:(1)The size of cervical cancer is very small,and existing segmentation methods are easy to neglect these small targets;(2)The size of tumor varies greatly in different slice of MRI,and the segmentation results of existing segmentation methods lack continuity and consistency;(3)Previous studies on cervical cancer tumor segmentation were few,resulting in a lack of labeled datasets.This paper focuses on the above issues,and the specific work is mainly divided into the following parts:(1)This paper proposes the Global-Local Cascaded Network(GLC-Net)for autonomously segmentating cervical cancer tumor from MRI,which can effectively solve the small target segmentation problem such as cervical cancer tumor.In order to effectively fuse multi-scale global features of images,Global Feature Fusion module is proposed;as local features are easy to lose in the process of down sampling and convolution,the Feature Decomposition Recombination module is proposed;the Channel Spatial Attention Module is designed to focus on specific targets.In order to demonstrate the segmentation effect of the proposed method,a series of comparison experiments and ablation experiments are designed.The results show that the method achieves satisfactory results in both quantitative indexes and visual effects.(2)This paper proposes the Slice Interaction Semi Medical Image Segmentation Network(SISMIS-Net),which only uses parts of labeled slices to train the tumor segmentation model.First use piece-wise affine transformation to enhance the data,and then train GLC-Net with parts of labeled slices to obtain a coarse segmentation model and get the coarse segmentation results of all slices.Then morphological transformation based on the slices interaction is used to transform the coarse segmentation results and generate soft labels.Finally,a semi supervised segmentation model based on spatial continuity is proposed to segment the coarse segmentation results more finely,supervised by soft labels or real labels.The effectiveness of the proposed method is verified by analyzing the results of comparison experiments and ablation experiments.(3)In order to further explore the clinical application value of medical image segmentation algorithm,we build a medical image segmentation system based on our proposed medical image segmentation model.The system provides a visual interface for the segmentation algorithm,including the function of user management,image upload and image segmentation.