Digital Image Analysis of Heterogeneous Tuberculosis Pulmonary Pathology in Non-Clinical Animal Models using Deep Convolutional Neural Networks

Abstract

Efforts to develop effective and safe drugs for treatment of tuberculosis require preclinical evaluation in animal models. Alongside efficacy testing of novel therapies, effects on pulmonary pathology and disease progression are monitored by using histopathology images from these infected animals. to compare the severity of disease across treatment cohorts, pathologists have historically assigned a semi-quantitative histopathology score that may be subjective in terms of their training, experience, and personal bias. Manual histopathology therefore has limitations regarding reproducibility between studies and pathologists, potentially masking successful treatments. This report describes a pathologist-assistive software tool that reduces these user limitations, while providing a rapid, quantitative scoring system for digital histopathology image analysis. The software, called 'Lesion Image Recognition and Analysis' (LIRA), employs convolutional neural networks to classify seven different pathology features, including three different lesion types from pulmonary tissues of the C3HeB/FeJ tuberculosis mouse model. LIRA was developed to improve the efficiency of histopathology analysis for mouse tuberculosis infection models, this approach has also broader applications to other disease models and tissues. the full source code and documentation is available from https://Github. com/TB-imaging/LIRA.

Comparative Analysis of Pharmacodynamics in the C3HeB/FeJ Mouse Tuberculosis Model for DprE1 Inhibitors TBA-7371, PBTZ169, and OPC-167832

“…(A) Blinded histopathology analysis by a board-certified pathologist using a semiquantitative method based on histology scores ( 56 ). (B) Quantitative analysis plus the SEM using pathologist-assistive software called LIRA (lesion image recognition and analysis [ 21 ]) developed for pulmonary pathology of M. tuberculosis -infected C3HeB/FeJ mice, employing convolutional neural networks trained using a machine learning approach, and (C) visual representation of the results generated by LIRA for digital images of pulmonary lesions of drug-treated C3HeB/FeJ mice. The various lesion classifications are displayed in a color overlay depicting type I core (red), type I rim (blue), type II (green), type III (yellow), healthy tissue (pink), miscellaneous (purple), and empty slide (no color).…”

“…Second, digital scans of hematoxylin and eosin (H&E)-stained tissue sections were evaluated using a new pathologist-assistive software called LIRA (lesion image recognition and analysis [ 21 ]), designed and developed for digital image analysis of pulmonary pathology in M. tuberculosis -infected C3HeB/FeJ mice. The LIRA results showed that all three DprE1 inhibitors were able to halt the progression of the lung pathology from treatment start by preventing the development of destructive type II neutrophilic lesions (which account for 25% of the total lung area in the untreated controls after 8 weeks) ( Fig.…”

“…The second histopathology method used a newly developed pathologist-assistive software called LIRA (lesion image recognition and analysis [ 21 ]), which reduces user limitations in terms of time and reproducibility while providing a rapid quantitative scoring system for digital histopathology image analysis. The LIRA source code with accompanying documentation has been made available at https://Github.com/TB-imaging/LIRA .…”

“…However, these mouse strains show some limitations by solely developing a single lesion type after M. tuberculosis infection and therefore lack the lesion heterogeneity seen in the lungs of TB patients ( 19 , 20 ). The C3HeB/FeJ mouse infection model presents with a heterogeneity of pulmonary TB lesion types more reflective of human disease ( 21 – 31 ) and is currently employed primarily in late-stage TB drug and regimen development ( 26 , 32 – 35 ). This mouse model was first described for tuberculosis by Kramnik et al and is commonly referred to as the “Kramnik mouse model” ( 36 ).…”

Subunit vaccine protects against a clinical isolate of Mycobacterium avium in wild type and immunocompromised mouse models

“…Percent lesion calculations were integrated into the same algorithm and calculated from tissue area and lesion area as designated by the ROI and lesions detected. Lesion identification and quantification were then reviewed and edited by a pathologist as needed 79 .…”