A Study On Medical Image Classification Algorithm Based On Meta-learning Framework

For medical images,the densely distributed microvascular regions in tissue section images are important features to judge the extent of tumor growth and malignancy.The classification algorithm for microvascular medical image classification tasks plays an im-portant auxiliary role in medical diagnosis.At the same time,skin cancer,as the most com-mon cancer in the world,is usually diagnosed by skin pigmentation and texture.However,because there are many complicated types of skin pathological and pathological charac-teristics,professional training is often required to complete the diagnosis.Therefore,an effective automated skin pathology detection algorithm can greatly alleviate the human re-sources required for the diagnosis of skin lesions.However,using deep neural networks to process image classification tasks often requires a large number of samples to participate in training.But in the medical field,it is difficult to meet the large amount of data required for deep learning.At the same time,deep learning network models based on large-scale data are weak in generalization and difficult to adapt to new classification tasks in a short time.Therefore,this thesis proposes to analyze the classification of medical images based on the meta-learning framework.The main research contents are as follows:1.The preprocessing methods of medical images are studied and analyzed.Based on the sensitivity of microvessel images to color features,this thesis uses a Single Oppo-nent(SO)feature extraction algorithm to preprocess the images.Aiming at the negative influence of the abnormal illumination of the skin pathological image on the characteris-tics of the pathological region,this thesis uses the Adaptive Surround Modulation(ASM)color normalization algorithm to correct the illumination.At the same time,due to the severely uneven number of skin pathological image samples,medical image data sets were expanded by combining image processing-based data augmentation and Conditional Gen-eration Adversarial Nets(CGAN).Then the maximum entropy method is used to filter out redundant samples,and finally the data is balanced.2.A meta-learning algorithm for medical image classification is researched and an-alyzed.Use the training classification samples of the meta-learner in Model-Agnostic Meta-Learning(MAML)to optimize the ring cross-correlation filtering template to per-form multiple batch voting on the prediction results of the classifier,and finally get the op-timized medicine Image classification results.At the same time,according to the MAML network structure,an improvement scheme is proposed,which improves the classification effect of the meta-learning classifier by adding convolutional layers,non-linear layers,and improved pooling methods.3.The Long Short-Term Memory(LSTM)model algorithm is researched and ana-lyzed,and an LSTM-optimized meta-learning algorithm is proposed for the classification of optimizing microvascular images based on the Double Opponent(DO)model.In the experimental part,the leading advantages of this method over other cutting-edge algo-rithms are introduced.In this thesis,a large number of experiments have been performed to verify that feature preprocessing algorithms which includes the sample equalization,SO-based,and ASM-based color normalization have improved the performance of medical image clas-sification algorithms based on meta-learning.An improved meta-learning algorithm for multi-batch voting using circular cross-correlation filter is proposed.And improve the classification performance by improving the meta-learning network.The proposed meta-learning algorithm based on LSTM optimization has achieved good classification results in the optimized microvascular image classification.In addition,while improving the classification effect of medical images,the meta-learner also exerts its generalization ad-vantage on different new tasks.This method of processing medical image classification provides a good theoretical and practical basis for the implementation of intelligent med-ical diagnosis.