Research On Pathology Image Detection Algorithm For Clinical Diagnosis

Pathological examination is the gold standard for the diagnosis of major diseases such as tumors,and pathologists complete pathological diagnosis and prognosis assessment by observing instances(cells and tissues)in pathological slides.The diagnosis process is a di?icult and intensive task that urgently requires intelligent analysis methods to aid diagnosis.Although various deep learning methods based on convolutional neural networks have been applied to a variety of detection tasks in recent years,but pathology images have different inherent characteristics from images of other types,there are still many problems to be solved in the process of extending pathology detection models from “laboratory research” to real clinical diagnosis.Specifically,according to the different levels of the task,this thesis distills and divides the key problems into three levels.(1)Di?iculties in training sample labeling: in the clinicalization process of detection models,a large amount of clinical data needs to be labeled to train the models,however,the high cost of pathology image labeling poses an obstacle to the extension of detection models;on the other hand,clinical data often have poor labeling quality,which reduces the training performance of detection models;(2)Di?iculties in the design of detection models: traditional convolution with only translational variability,is di?icult to extract features of pathological instances with variable angles;meanwhile,the traditional detection architecture is often susceptible to interference from the complex background of pathological images;(3)Di?iculties in the incorporation of medical a priori knowledge in clinical diagnosis: The connotation of the diagnostic task of pathology image detection goes far beyond the simple image analysis task,and the detection model needs to be designed in a way that fits the clinical process and specifications,if this requirement can be achieved,not only can the diagnostic performance of the model be improved,but also the automatic diagnostic results can be more easily accepted by pathologists.However,how to integrate deep learning paradigms with clinical knowledge remains to be studied.In this thesis,we conduct research and innovation based on the above-mentioned di?iculties,hoping to promote more effective implementation of pathology image detection algorithms in clinical diagnostic tasks and improve the reliability and interpretability of diagnostic results.The results achieved are as follows.1.Discriminator-based high-value sample selection method.The distribution of pathological images from different institutions shows heterogeneity due to factors such as stains and scanner parameters,resulting in the same model can not be accurately adapted to multiple institutions.Therefore,before deploying the model to a new institution,training data from the corresponding institution should be constructed to iterate on the model,and the cost of data labeling accounts for a major part of the iteration cost.To reduce the annotation cost,samples with high value should be selected for annotation.Based on this,this thesis proposes an active learning method for pathology images,which uses a deep discriminator to unbiasedly estimate the instance values in pathology images,and thus selects the most valuable samples for annotation to iterate the model,effectively reducing the data cost.2.Robust training strategy for pathology instance detection.The training samples of clinical detection models need to be annotated at the instance level by professional pathologists.However,the dense distribution of instances in pathology images makes full annotation is time-consuming and not easy to implement,so the trainning process often faces data with poor annotation quality,which causes di?iculties in model training.In this thesis,we propose a loss calibration strategy based on energy density and a feature mining strategy for spatial cues,and experimental results show that the model trained by the proposed method has stable performance even if when 90% of annotations are missing.3.Rotating covariant feature extraction convolutional kernel.Traditional convolution cannot effectively extract key features with variable angles in pathological instances.An orthogonal vector convolution kernel is proposed,which extends the translation invariance of traditional convolution to variability such as the Euclidean group transform,effectively enhances the feature capacity of convolution kernel parameters,constructs a more reliable instance key feature extractor,and lays the foundation for realizing a more accurate pathological instance detection model.4.Instance feature complementary architecture.Faced with the complex background structure of digital pathology images,the corresponding image analysis work becomes much more di?icult.In this work,we propose a novel deep learning detection architecture to mitigate the interference of complex backgrounds in pathology images.The architecture automatically stores global key features and continuously adds effective features to the instance regions during the decoding process to mitigate the influence of complex microstructures,which greatly eliminates the false detection of pathology image backgrounds.5.HER2 automatic interpretation scoring algorithm.An automated scoring algorithm for immunostaining HER2 slides that incorporates rules for section cancer area constraints and fits into the clinical diagnostic process is proposed.The slides automatic diagnosis results showed that the design of the detection algorithm according to the clinical process-fit approach can significantly improve the interpretation of pathological slides.