Research On Prostate Image Aided Labeling Technology Based On Deep Learning

In recent years,the combination of deep learning and Magnetic Resonance Imaging(MRI)to assist in the detection of prostate cancer has become a popular direction.The quality of deep learning model training relies on a large amount of high-quality annotated data.However,manual annotation of prostate MRI images requires a high level of medical expertise,and the annotation process is extremely time-consuming and laborious.Therefore,this article proposes an interactive image annotation framework based on composite geodesic distance,starting from the characteristics of prostate MRI images,and annotating the corresponding prostate tissues through segmentation.The main contents include:(1)By collaborating with a tertiary hospital,image data of 430 prostate cancer patients were collected and underwent a series of preprocessing operations such as image slicing,cropping,and enhancement.(2)This thesis proposes an interactive image annotation framework based on composite geodesic distance.The framework includes two stages: coarse segmentation and fine segmentation.In the coarse segmentation stage,the U-Net network is used as the backbone,and a conditional random field module is designed as post-processing to address the feature loss problem.A soft attention mechanism is also introduced to enhance the feature extraction of the prostate segmentation area and reduce interference from irrelevant information.To address the issue of imbalanced positive and negative sample distribution in prostate MRI images,a Cross-Dice Loss function is designed by combining cross-entropy Loss and Dice Loss to improve segmentation performance.In the fine segmentation stage,user interaction information is incorporated using the composite geodesic distance algorithm based on the coarse segmentation to optimize the segmentation.Finally,this paper trains and tests the algorithm on 120 annotated prostate cancer patient cases.Through comparative experiments,the algorithm in this article achieved a Dice score of 93.67% in prostate region segmentation,which is superior to other segmentation algorithms.(3)A medical image annotation prototype tool based on interactive segmentation was designed and implemented.The prototype tool was validated through application,and was used to interactively annotate the collected unannotated data of this thesis,resulting in a high-quality prostate annotated image dataset with 430 cases that is accepted by doctors.