Intelligent Analysis Algorithm Research On Digital Pathology Images Based On Deep Learning

Pathological diagnosis,as a kind of medical diagnosis which depends on direct inspection on the histopathological structures and physiological characteristics of the human biopsy samples under an microscope,is the authoritative diagnosis method in medical diagnosis domain.It is also known as the "gold standard".Centered on the intelligent analysis of digital pathology images,the research described in this paper mainly contains the segmentation of basement membrane in micro-invasive cervical carcinoma,nuclear segmentation,and interpretable diagnosis algorithm combining the semantic information latent in the pathology diagnosis reports:1)Existing semantic segmentation or semantic contour segmentation models have the following issues when directly applying to the segmentation of basement membrane in micro-invasive cervical carcinoma: low integrity,discontinuity and low globality.To address these issues,we proposed a semantic contour segmentation model based on adversarial networks,to teach the segmentation networks to learn the high-level pixel consistency and the shape priors of the basement membrane.Extensive experimemts on the dataset built by the pathologists demonstrated that the proposed model has solved the issues described above well2)Accurate segmentation of nuclei in digital pathology images is faced with the following challenges: severe overlapping,inner-nuclei intensity inhomogeneity,complex artifacts in noisy backgrounds,technical variations like dye differences,etc.We proposed a cascaded convolutional network model to perform four sub-tasks: the nuclei foreground extraction,nuclei foreground denoising,nuclei foreground distance transform and nuclei boundary extraction,which are integrated into one end-to-end trainable framework.Extensive experiments demonstrate the superiority of the proposed model comparing to the exisiting methods.3)Most existing cancer grading or recognition research regard these problems as simple image classification tasks.However,in practice,pathologists often write diagnosis reports to describe what they see in the image so that they can explain why they infer the following diagnosis conclusion.Based on this fact,we aim to do research on how to effectively utilize the semantic information latent in the pathological reports to facilitate the diagnosis conclusion prediction.Besides,based on the attention mechanism in the natural image research field,we further introduce interpretability into our model,in the way of visualizing attentional regions on the input image.Based on the dataset(with both images and reports)built by pathologists,we compared our model with existing classification models on the classification accuracy.Extensive experiments results demonstrate that the introduced semantic knowledge has positive impact on the diagnosis conclusion prediction.