Research On Morphological Lesion Recognition And Segmentation Methods In Glomerular Pathology Imaging

Chronic kidney disease(CKD)is a global public health issue of great concern.Its diagnosis and treatment efficiency are highly correlated with the socioeconomic burden it causes.Intelligent assessment of human renal biopsies could assist physicians in achieving efficient diagnosis and treatment.Identification and segmentation of glomerular morphological lesions is a critical step in the histopathologic evaluation of human renal biopsies.In this paper,we focus on the recognition and segmentation method of three typical glomerular morphological lesions in renal pathology images: sclerosis,segmental sclerosis and crescent.The main work includes dataset construction,network development and experimental verification as follows:(1)Under the guidance of professional doctors from the Department of Renal Pathology at Sichuan Provincial People’s Hospital,we collect renal biopsy whole slide images stained with periodic acid-Schiff(PAS),periodic acid-silver metheramine(PASM),and Masson’s trichrome(Masson),and label the data containing one type of morphological lesion.The first dataset for identifying and segmenting glomerular morphological lesions was constructed to provide data support for subsequent studies.(2)To address the problem of high inter-class texture and scale similarity and high intra-class morphological differentiation of glomerular lesions,we propose a recognition network with enhanced discriminable information for identifying glomerular morphological lesions.The method first improves the semantic feature distinction of lesions by enriching and enhancing the middle and high level semantic discriminable features of pathological images.Then,the recognition performance of lesions is further improved by a multi-level self-attention mechanism,which perceives the relative position relationship between the lesion and the glomerulus and the scale of the lesion.The experimental results shows that the recognition accuracy of this method for three types of typical glomerular morphological lesions reaches over 92.00%.(3)For the problem of small morphological and textual differences in glomerular lesions and blurred lesion boundaries in stained pathological images,a context-memory network for glomerular morphological lesions is proposed.This method first designs the ConvLSTM-ASPP module to extract the advanced semantic multiscale features of the lesions,and to clarify the lesion scale and the relative position relationship between the lesions and the glomerulus.Then,ConvLSTM is used to integrate the multi-level features of the lesion,which makes the target texture features match with the high-level semantic features and sharpens the contour of the lesion.This method realizes the accurate segmentation of the glomerular morphological lesion.The experimental results show that performance of context-memory network for three morphological lesions is 67.91%mIoU.Based on the identification and segmentation datasets of glomerular morphological lesions,experimental verification demonstrates the methods we proposed for glomerular morphological lesions.