Research On Key Technologies Of Intelligent Diagnosis By Medical Magnetic Resonance Imaging And Optical Microscopic Imaging Equipment

The research on intelligent diagnosis technology of medical images and microscopic images belongs to the interdisciplinary fields of optical engineering,computer science,and medicine.Due to its excellent performance in alleviating the workload of doctors,reducing the gap in medical resources,and early detection of hidden diseases,it is gradually becoming an important part of our country and a new trend of medical development in the future.However,there are two major bottleneck problems in intelligent diagnosis technology research: "weak generalization ability of models under small dataset scenarios" and "poor feature learning ability of models under complex scenarios".In the fields of medical imaging and microscopy,these problems manifest as many challenges,such as scarce sample size,poor consistency,limited annotation,and difficult multi-object detection.Therefore,this dissertation conducts comprehensive research on four subjects,including qualitative diagnosis of prostate cancer MRI(Magnetic Resonance Imaging),qualitative diagnosis of brain tumors MRI,qualitative diagnosis of vaginitis,and quantitative detection of vaginal microorganisms.And this dissertation proposes a series of methods to improve the model’s generalization ability and feature learning ability,ultimately achieving the goal of improving the accuracy and efficiency of intelligent disease diagnosis.The main research results are as follows:1.An intelligent classification algorithm based on channel-independent random rotation data enhancement is proposed,which enables intelligent diagnosis of MRI with scarce data from a single institution.Medical MRI is characterized by high acquisition costs,small datasets,low resolution,severe noise,and a small proportion of lesions,which can significantly hinder model classification performance.A shallow convolutional neural network named Pca-Shallow Net and a channel-independent random rotation data augmentation algorithm are designed.The proposed Pca-Shallow Net has the characteristics of small convolution kernel size,zero padding and mixed dropout,which effectively improves the classification performance.At the same time,the channelindependent random rotation data augmentation method introduces the prior knowledge of experts,and then estimates the neighborhood around each sample in the original training set based on the principle of vicinal risk minimization,so as to realize the empirical distribution estimation of the image data set.It significantly expands the dataset and improves the generalization ability of the model.On the prostate cancer MRI dataset,the classification AUC(the Area Under the ROC Curve)is 85.04%,which is 4.81%higher than the baseline.The proposed algorithm effectively reduces the damage caused by the scarcity of sample size to the model’s performance.2.An intelligent grading algorithm based on transfer learning and multi-scale feature fusion is proposed,which enables intelligent diagnosis of MRI with inconsistent samples from multiple institutions.When medical MRI is collected from devices of different brands and models,differences in data resolution,noise,and artifact size can lead to inaccurate feature extraction,while the limited dataset can also hinder the model’s learning ability.A block-wise mode transfer learning method and a multi-scale feature fusion model architecture named Glioma-AFPNet are designed.The proposed transfer learning method can maximize the knowledge learned on large datasets to target datasets,and the proposed model is designed based on a lightweight channel attention mechanism and a feature pyramid network,which weights the channel feature maps of each convolutional block,and then combines the low-dimensional features which are rich in spatial information(shape,edge,etc.)and high-dimensional feature maps which are rich in semantic information(position,contour,etc.).The proposed algorithm strengthens and improves the feature extraction and learning capabilities.On the brain tumor MRI dataset,the grading AUC value is 90.42%,which is 10.24% higher than the baseline.The proposed algorithm effectively solves the problem of poor data consistency.3.An intelligent classification algorithm based on multi-network ensemble active learning is proposed,which enables intelligent diagnosis of microscopic images for qualitative assessment of medical images with high annotation costs.For medical microscopic images with complex backgrounds and multiple cell types,the high annotation time and manpower costs,coupled with limited annotation quantities,can lead to poor classification accuracy of the model.An active learning framework named LeuECNNs that integrates multiple lightweight network architectures is proposed,which utilizes multiple network characteristics,such as changes in depth,convolution filter size,residual mechanism,and attention mechanism to improve the feature extraction and learning ability for complex background microscopic images.Based on Bayesian convolutional neural network uncertainty sampling and committee querying strategies,the framework continuously selects the most valuable subset of training samples from the unlabeled sample pool for model training.Finally,a soft voting mechanism is employed for multiple network ensemble learning for inference prediction.The proposed framework improves the model’s generalization ability and microscopic image classification accuracy under low annotation data volume.On the vaginitis leucorrhea microscopy dataset,it saves 74.1% of the annotation volume and achieves 95.38%accuracy,96.36% precision,92.98% recall,and 94.64% F1 score.The proposed framework provides a feasible solution for the problem of limited annotation.4.An intelligent detection algorithm of visible components in complex backgrounds based on the attention mechanism is proposed,which enables quantitative analysis and intelligent diagnosis of formed elements in complex backgrounds.Due to the large variety and imbalanced distribution of formed elements in medical microscopic images,and the common occurrence of overlapping and adhesion,traditional object detection models struggle to balance detection speed and accuracy.In this dissertation,a framework based on the attention mechanism named Leu-DETR is proposed,which integrates a weighted random sampling module with dynamic data augmentation,introduces the grouping convolution mechanism and the relative position encoding module to enhance the feature extraction and learning ability of the model for overlapping and adhesion of multiple objects.It also utilizes multi-head attention mechanism and bipartite graph matching process to improve the detection accuracy of formed elements in complex backgrounds and effectively reduce detection time.On the vaginal microbiota microscopy dataset,the proposed algorithm achieves 86.76% m AP(mean Average Precision)for formed element detection,which is 1.89% higher than the baseline,and the AP(Average Precision)of the minority class cells is 4.65% higher than the baseline.The detection time per frame is about 78.4 milliseconds.The proposed algorithm has achieved satisfactory results in addressing the problem of difficult multi-object detection.In summary,this dissertation takes the bottleneck problems in the field of intelligent diagnosis technology as guidance,conducts in-depth research,and achieves certain results,which can effectively promote the theoretical development and practical clinical application of intelligent diagnosis of medical imaging and microscopic imaging equipment.