Research On Key Technologies For Medical Image Processing Based On Deep Learning With Imperfect Datasets

The medical image processing based on deep learning has gradually become a pow-erful tool for assisting doctors in clinical practice due to its high efficiency,accuracy and objectivity.However,the training of deep learning models often requires a large number of precisely labeled training samples,which brings great challenges to the application of deep learning in medical image processing.Because the precise annotation of medical im-ages often requires experienced clinicians to perform,the acquisition cost of large datasets is often too expensive.To build medical image datasets for training deep learning models,researchers try to download training samples from the previous diagnostic cases of hospi-tals,and extract sample labels manually or based on natural language processing from the diagnostic report corresponding to the sample.The dataset constructed based on the above method contains constraints such as noisy labels,category ambiguity,small samples,and domain drift that degrade the generalization of deep neural networks.Therefore,the re-search on medical image processing based on deep learning with imperfect datasets has attracted the attention of more and more researchers,and has become a research hotspot in the field of medical image processing.This dissertation mainly focuses on the key challenges of deep learning for the com-mon sample constraints in lung image analysis task,with the overall goal of building robust deep medical image processing models,and provides novel deep learning theo-ries and methods that are robust to sample constraints in medical image processing.In detail,this dissertation conducts research on robust segmentation architecture,noisy sam-ple selection based on multi-network collaborative learning,noisy sample mining based on self-supervised learning and prediction consistency loss,ambiguous sample learning based on robust loss functions,and ambiguous sample relaxation based on robust regular-ization.The main contributions are as follows:1.To address the low generalization limitation of deep segmentation models caused by sample constraints such as noisy labels,category ambiguity,small samples,and do-main drift between different datasets,this dissertation proposes an asymmetric encoder-decoder lung field segmentation network for chest radiographs.The network first learns the common lesion representation between multi-source samples to reduce the domain drift between multi-datasets,and then uses the sample resampling to reduce the number of pixels with noisy labels and category ambiguity in the training samples,and finally re-duces the model’s overfitting to small samples by embedding an attention mechanism in the encoder-decoder network architecture.The network is trained and validated on JSRT lung nodule and Montgomery tuberculosis hybrid datasets,and achieves lung field seg-mentation that is more robust than the classical FCN and U-Net on the pneumoconiosis staging dataset.2.The multi-network collaborative learning methods based on the small loss strat-egy can filter out the noisy samples in the training set,but when the training samples are mixed with category ambiguity and the recognition difficulty of each category is in-consistent,this type of methods will filter out accurately labeled ambiguous samples and difficult-to-recognize category samples,resulting in a deep learning model that can only fit the local representation of the training samples.Aiming at the above problems,this dissertation proposes a pneumoconiosis chest radiograph classification method based on multi-network collaborative learning inspired by the clinical expert consultation mecha-nism.The method first uses a triple-network architecture to simulate an clinical expert consultation group,and then divides the samples in the training set into three subsets of ‘clean-easy’,‘clean-hard’ and ‘noisy-hard’ through the quaternary sample partition-ing strategy,and then performs relaxed learning on samples in the ‘clean-hard’ subset by loss re-weighting,and finally mines sample features in the ‘noisy-hard’ subset through a self-supervised learning strategy based on data augmentation and prediction consistency loss.Experimental results prove that this method overcomes the problem of insufficient filtering ability of existing methods based on multi-network collaborative learning,and re-alizes chest radiograph classification of pneumoconiosis that is more robust than classical methods such as Decoupling,Co-teaching and Jo Co R.3.Label distribution learning can alleviate the model overfitting caused by category ambiguity and noisy labels in the training set through estimating the conditional class prob-ability distribution of each category.However,the label distribution constructed based on the normal distribution does not match the asymmetric characteristics of the ambiguity and noise in the pneumoconiosis staging dataset,and the label distribution learning based on the KL-divergence loss is likely to lead to subjective inconsistencies.Aiming at the above problems,this dissertation proposes a pneumoconiosis chest radiograph classifica-tion method based on log-normal label distribution learning.This method first samples the asymmetric label distribution of each category of pneumoconiosis based on the log-normal distribution to replace the original given one-hot label,and then introduces cross-entropy loss as a regularization term in the training loss to solve the subjective inconsistency of KL-divergence loss,and finally designs a dropout parameter to solve the optimization ob-jective inconsistency between KL-divergence loss and cross-entropy loss.Experiments demonstrate that this method improves the robustness of Res Net to category ambiguity and noisy labels in the pneumoconiosis staging dataset.4.The robust regularization method based on label smoothing greatly alleviates the overfitting of deep neural networks to the training samples by introducing a uniform dis-tribution in the labels.However,the label-smoothed uniform probability distribution does not match the true conditional class probability,which introduces biases that degrade model generalization during the training of deep neural networks.The label relaxation strategy can make up for the shortcomings of label smoothing,but the instance-level es-timation of the exact sample relaxation degree is a huge challenge.To address the above issues,this dissertation proposes a dynamic label relaxation strategy to estimate the degree of sample relaxation,and applies this method to the lung image analysis task.This method first measures the ambiguity of clean samples based on the average prediction entropy of the multi-backbone network,and then sorts the clean samples according to the average prediction entropy,and finally samples the hyper-parameter of the deep neural network for relaxation learning for each clean sample based on a cosine annealing schedule or an exponential descent schedule.Experiments demonstrate that the proposed dynamic label relaxation strategy improves the generalization of deep neural networks on the pneumo-coniosis staging dataset and the UESTC-COVID-19 pneumonia segmentation dataset.In short,this dissertation proposes the above four lung image analysis methods,ex-plores the key technologies of lung image processing based on deep learning under sample constraints,and proves that deep learning methods are still feasible in medical image pro-cessing related fields with sample constraints.