Deep learning enables pathologist-like scoring of NASH models

Abstract

Non-alcoholic fatty liver disease (NAFLD) and the progressive form of non-alcoholic steatohepatitis (NASH) are diseases of major importance with a high unmet medical need. Efficacy studies on novel compounds to treat NAFLD/NASH using disease models are frequently evaluated using established histological feature scores on ballooning, inflammation, steatosis and fibrosis. These features are assessed by a trained pathologist using microscopy and assigned discrete scores. We demonstrate how to automate these scores with convolutional neural networks (CNNs). Whole slide images of stained liver sections are analyzed using two different scales with four CNNs, each specialized for one of four histopathological features. A continuous value is obtained to quantify the extent of each feature, which can be used directly to provide a high resolution readout. In addition, the continuous values can be mapped to obtain the established discrete pathologist-like scores. The automated deep learning-based scores show good agreement with the trainer - a human pathologist.

Potentials of AI in medical image analysis in Gastroenterology and Hepatology

“…A large gastric histopathology dataset with pixel‐level annotations was constructed, and they achieved an accuracy outperforming conventional methods. Heinemann et al 28 . automated histological feature (ballooning, inflammation, steatosis, and fibrosis) scoring for non‐alcoholic steatohepatitis (NASH) with DL methods.…”

“…A large gastric histopathology dataset with pixel-level annotations was constructed, and they achieved an accuracy outperforming conventional methods. Heinemann et al 28 automated histological feature (ballooning, inflammation, steatosis, and fibrosis) scoring for non-alcoholic steatohepatitis (NASH) with DL methods. The DL algorithm output continuous scores for quantifying the extent of each feature and the quantitative comparison with human pathologists achieved very good agreement, such as steatosis.…”

Molecular Imaging of Inflammatory Disease

“…To give some idea of these performance values, the range of the Dice coefficient for sinusitis segmentation is 86%-97%, the accuracy is 78% to 92% [190,191,204]. For COPD and lung disease inflammation, and the accuracy ranges from 61% to 95% [177,192,193,195,205]. 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].…”

“…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.…”

A Large-Scale Internal Validation Study of Unsupervised Virtual Trichrome Staining Technologies on Non-alcoholic Steatohepatitis Liver Biopsies

“…These steps can facilitate faster clinical decision making with less resources (both departmental and tissue). Future analyses may explore the prediction of Fibrosis stage from an H&E, obviating the need for clinical inspection of a virtual stain 68 . However, virtual staining is a technology that clinicians may trust more than a computer-generated risk score and therefore presents an immediately viable decision aid technology.…”