Deep neural network based histological scoring of lung fibrosis and inflammation in the mouse model system

Abstract

Preclinical studies of novel compounds rely on quantitative readouts from animal models. Frequently employed readouts from histopathological tissue scoring are time consuming, require highly specialized staff and are subject to inherent variability. Recent advances in deep convolutional neural networks (CNN) now allow automating such scoring tasks. Here, we demonstrate this for the case of the Ashcroft fibrosis score and a newly developed inflammation score to characterize fibrotic and inflammatory lung diseases. Sections of lung tissue from mice exhibiting a wide range of fibrotic and inflammatory states were stained with Masson trichrome. Whole slide scans using a 20x objective were acquired and cut into smaller tiles of 512x512 pixels. The tiles were subsequently classified by specialized CNNs, either an “Ashcroft fibrosis CNN” or an “inflammation CNN”. For the Ashcroft fibrosis score the CNN was fine-tuned by using 14000 labelled tiles. For the inflammation score the CNN was trained with 3500 labelled tiles. After training, the Ashcroft fibrosis CNN achieved an accuracy of 79.5% and the inflammation CNN an accuracy of 80.0%. An error analysis revealed that misclassifications are almost exclusively with neighboring scores, which reflects the inherent ambiguity of parts of the data. The variability between two experts was found to be larger than the variability between the CNN classifications and the ground truth. The CNN generated Ashcroft score was in very good agreement with the score of a pathologist (r2 = 0.92). Our results demonstrate that costly and time consuming scoring tasks can be automated and standardized with deep learning. New scores such as the inflammation score can be easily developed with the approach presented here.

Deep learning enables pathologist-like scoring of NASH models

“…In all four models, we used an “ignore” class to identify all cases where liver tissue was not-sufficiently visible on a tile or other artifact types were present (e.g. out of focus, mostly blood, or staining artifacts), as in our previous work 25 . The “ignore” class ensured, that only tiles containing actual liver tissue were further analyzed.

Figure 2Examples of the classes used to train the four convolutional neural networks (CNN) to recognize relevant features of the histopathological features in the Kleiner score.

…”

“…If the loss on the validation data did not decrease for more than two epochs, the learning rate was reduced by multiplying with a factor of 0.2 to a minimal learning rate of η = 10 −7 . No further hyperparameter tuning was performed since previous experience with these parameters and Inception-V3 resulted in very good recognition performances with higher agreement levels in tile recognition compared to two human experts 25 . Class imbalances were equalized by oversampling.…”

“…The number of applications in histopathology is rapidly growing. Examples include prostate cancer diagnosis 24 , non-small cell lung cancer diagnosis and mutation prediction 11 or pathologist-like fibrosis and inflammation scoring in lung tissue 25 .…”

Molecular Imaging of Inflammatory Disease

“…Some studies go beyond detection and segmentation to scoring severity ( Figure 7) [173,189]. Computerized analysis of inflammation was also applied to other inflammation diseases: paranasal sinus [190,191], chronic obstructive pulmonary disease (COPD) [177,[192][193][194], celiac disease (CD) [195][196][197][198], inflammatory gastrointestinal lesions [176,199,200], varicose veins [201], myocarditis [202,203], and inflammatory brain abnormalities. 7.…”

“…For CD, the accuracy ranges from 79% to 97%, while the sensitivity and specificity vary from 83% to 100% and 96% to 100%, respectively [195][196][197][198]. [177,[192][193][194]205] Celiac Disease (CD) Endoscopy Images H&E Duodenal Biopsy Images CNN-Based Transfer Learning (Alexnet, VGG Nets, Resnet) SVM, Bayesian [195][196][197][198] Inflammatory Many reported studies need larger training datasets to better characterize the bias among different imaging modalities and to improve their performance and generalizability because of the variability in datasets. These studies also highlight the need for stronger clinical significance.…”