The Research Of Deep Learning Algorithms On Classification And Detection Of Pathological Images Under The Condition Of Data Inconsistencies

Applying deep learning to pathological image classification and detection is one of the fundamental researches in intelligent pathological diagnosis.Deep learning algorithms can assist pathologists in diagnosing,which save their time and obtain higher diagnostic accuracy than professional pathologists under complex conditions.However,the effectiveness of deep learning algorithms heavily depends on the consistency of training data and test data.Data inconsistency can affect the performance of deep learning algorithms.The thesis focuses on the background of intelligent pathological diagnosis of gastric cancer and explores how to improve the accuracy of deep learning algorithms for pathological images classification and signet ring cells detection under conditions of color difference,resolution difference,and incomplete labels.For the color difference in the dataset,based on the structure-preserving color normalization(SPCN)algorithm,the K-means and SNMF basis combination(KSBC)algorithm is proposed in the thesis by combining the K-means cluster algorithm and sparse non-negative matrix factorization(SNMF)model.It inherits the structure-preserving feature of the SPCN algorithm,and can flexibly migrate the color of pathological images.Furthermore,a deep learning algorithm for classification is proposed based on the preprocessing of KSBC.Experiments show that it can improve the classification accuracy of pathological images.For the difference in resolution between the training dataset and the test dataset,and incomplete labels in the training dataset,the thesis combines a super-resolution algorithm,i.e.USRNet(unfolding super-resolution network),and innovatively designs a revised gradient harmonizing mechanism classification(RGHMC)loss function by the label correction mechanism,a deep learning-based algorithm for signet ring cell detection REUR(Retina Net Embedded into USRNet with RGHMC loss)is proposed in the thesis.The test results on the gastric dataset of Nanfang Hospital show that REUR in the thesis can accurately detect signet ring cells from actual low-resolution pathological images.In general,the thesis focuses on the difficulties caused by data inconsistencies in gastric pathological image classification and signet ring cells detection for deep learning algorithms and studies the algorithm design from the data preprocessing,deep neural networks,and loss functions.The proposed new algorithms can provide support for diagnosing gastric pathological images,which are also helpful for other medical image analysis.