Feature Learning And Representation Of Medical Histopathology Images

Medical image screening is recognized as the gold standard of clinical diagnosis,and its accuracy and speed are of vital importance to precision medicine in the future.However,medical images usually have a high resolution with rich information of tissue structure and cell morphology,and the lesions are highly heterogeneous,leading to a time burden for the diagnosis of a single image.The scarcity of professional doctors further affects the efficiency of treatment.Thanks to the remarkable achievements of deep convolutional neural networks in natural image processing,histopathological image analysis based on transfer learning is burgeoning.This paper focuses on the challenging task of cervical histopathological image analysis,so as to lay a foundation for the practical application of intelligent medical treatment.The contributions can be summarized as follows:1.Faced with the scarcity of high-quality datasets,a cervical precancerous histopathology image dataset called MTCHI(Multi-Task Cervical Histopathological Images)is constructed.The MTCHI dataset consists of 100 gigapixel resolution slides from 71 patients and is annotated independently by three pathologists.Four evaluation metrics are provided for fair comparison of algorithms.In addition,strong baselines are provided by conducting massive experiments derived from classification and segmentation networks,which demonstrates the feasibility of computer-aided diagnosis based on deep convolutional neural networks and the room for improvement.2.By exploring the relationship between histopathological categories,three loss functions are proposed to optimize the intelligent analysis of breast and cervical images.First,a gravitation loss is designed to improve the feature extraction capability of the model by reducing the intra-class variance and increasing the inter-class distance.Second,a distribution consistency loss is proposed according to the relationship between cervical classes.Third,an adaptive elastic loss is proposed to draw a difficult sample to the correct category by assigning a dynamic penalty weight to the misjudgment.Experimental results on public datasets demonstrate the validity of the gravitation loss,the distribution consistency loss and the adaptive elastic loss.3.Two end-to-end network architectures for exploiting the segmentation potential are proposed.First,TriUpSegNet(Triple Upsampling Segmentation Network)is constructed to enhance the attention of the center area of the image,and a Gauss-like weighted post-processing method is proposed to suppress the edge noise.Second,considering the advantages of the typical segmentation networks,a HSP-Net(Hierarchical Spatial Pyramid Network)with strong spatial feature expression ability is reconstructed.The ablation experiments and cross validation on the MTCHI dataset reflect the superior performance of TriUpSegNet and HSP-Net and the potential of segmentation methods for intelligent analysis.4.A dual-task collaborative network that can be trained with image-level annotations is proposed.First,a dual-task collaborative network is assembled by integrating the classification and segmentation networks,in which the global information is controlled by the classification branch and the local information is controlled by the segmentation branch.Second,a weakly supervised learning strategy based on pseudo-labeling is proposed.Finally,fully supervised and weakly supervised learning strategies are aggregated for the dual-task collaborative network.It is superior in speed and accuracy in the MTCHI dataset,and provides a direction for the training of a large number of auxiliary data in future clinical applications.