Research On Intelligent Screening Method For Prostate Cancer Ultrasound Images Based On Deep Learning

Prostate cancer is a major global public health problem and has become one of the most prevalent cancers in men.To improve the treatment rate and the quality of life of patients,prostate cancer screening is often used to detect the cancer at an early and treatable stage.However,the existing screening methods for prostate cancer,such as rectal examination and prostate-specific antigen blood test,have high uncertainty and low sensitivity and specificity.On the other hand,puncture biopsy,which is the gold standard for prostate cancer diagnosis,is an invasive screening modality that has a risk of infection and has greater side effects.Therefore,it is necessary to seek reliable and accurate automated screening methods.Since transrectal ultrasound image screening has the advantages of being noninvasive and economical,it is of great clinical value to study automatic screening methods for prostate cancer based on ultrasound images.The advantages of deep learning for natural image processing make its application in the field of medical image analysis attract much attention.However,the following challenges remain:(1)Ultrasound images have inherent noise,low contrast and low signal-to-noise ratio,and the existence of complex backgrounds interferes with the accuracy of prostate cancer screening to some extent;(2)High interclass similarity and large intraclass variability.Ultrasound images have low contrast and unclear details,and images of the prostate in the benign category have similar grayscale distributions to those in the malignant category,and the images have large differences even in the same category due to uncertainties in the ultrasound image acquisition process;(3)The sample size of prostate ultrasound images is relatively small,and the model tends to treat individual samples with unique properties as general properties to model,resulting in poor generalization of the model.In view of the above challenging problems and the characteristics of prostate ultrasound images,this thesis investigates prostate cancer ultrasound image screening based on deep learning techniques,and proposes a prostate gland segmentation method based on region labeling network and a classification method based on embedding augmentation metric learning,respectively.The details are as follows:(1)In order to mask the prostate background interference information to improve the accuracy of prostate cancer screening,this thesis proposes the Region Labeling Object Detection Network(RLOD)for achieving automatic segmentation of the prostate.This network first extracts the candidate regions of the prostate,then inputs the candidate region features into two branches for localization and segmentation,which finally obtains the labeling results of the prostate region.In RLOD,the Skip-connected Feature Pyramid Network(CFPN)module is proposed to enhance the fusion of high-level features and low-level features to improve the ability of the network to extract image detail information.Moreover,an attention mechanism is introduced in CFPN to enable the model to focus its attention on useful information.The proposed RLOD shows excellent segmentation performance on the prostate ultrasound dataset collected from hospitals,with an average DICE metric of 0.935,which demonstrates its effectiveness in prostate segmentation.(2)In view of the problems of low differentiation and small sample size of prostate cancer ultrasound images,this thesis proposes an Embedding Augmentation Metric Learning Network(EAML)to generate differentiated feature representations of prostate regions,and thus build an intelligent prostate cancer classifier based on differentiable features.For EAML,the network first uses a convolutional neural network to extract image features and map them to a lowdimensional embedding space,and introduces a Linear Interpolation Enhancement Strategy to generate new embeddings based on this space.Then,EAML remaps the synthetic embeddings based on the proposed Permutation-aided Reconstruction Loss to obtain synthetic samples for expanding the original samples.Finally,EAML uses the proposed Semantic Differences Mining Strategy to address the challenges of inter-class and intra-class differences based on the original and generated samples.The experimental results show that the EAML-based classifier performed high classification accuracy on the prostate regional ultrasound image dataset,with benign and malignant classification precision of 0.857 and 0.889,respectively,which have higher accuracy compared to the manual diagnosis by ultrasonographers.This suggests that the proposed EAML network can improve the screening of prostate cancer based on ultrasound images.