Using Artificial Intelligence to Detect, Classify, and Objectively Score Severity of Rodent Cardiomyopathy

Abstract

Rodent progressive cardiomyopathy (PCM) encompasses a constellation of microscopic findings commonly seen as a spontaneous background change in rat and mouse hearts. Primary histologic features of PCM include varying degrees of cardiomyocyte degeneration/necrosis, mononuclear cell infiltration, and fibrosis. Mineralization can also occur. Cardiotoxicity may increase the incidence and severity of PCM, and toxicity-related morphologic changes can overlap with those of PCM. Consequently, sensitive and consistent detection and quantification of PCM features are needed to help differentiate spontaneous from test article-related findings. To address this, we developed a computer-assisted image analysis algorithm, facilitated by a fully convolutional network deep learning technique, to detect and quantify the microscopic features of PCM (degeneration/necrosis, fibrosis, mononuclear cell infiltration, mineralization) in rat heart histologic sections. The trained algorithm achieved high values for accuracy, intersection over union, and dice coefficient for each feature. Further, there was a strong positive correlation between the percentage area of the heart predicted to have PCM lesions by the algorithm and the median severity grade assigned by a panel of veterinary toxicologic pathologists following light microscopic evaluation. By providing objective and sensitive quantification of the microscopic features of PCM, deep learning algorithms could assist pathologists in discerning cardiotoxicity-associated changes.

Keywords: Sprague Dawley; artificial intelligence; cardiomyopathy; computer assisted image analysis; computer neural networks; deep learning; rat.

Figure 1.

Schematic of the algorithm analysis and output pipelines. Input data consisted of WSI subdivided into tiles. In the upper pipeline (abnormality pipeline), a deep neural network model was trained on the tiles to identify PCM-related abnormalities of necrosis, fibrosis, MNC, and mineralization. In the lower pipeline (artifact pipeline), a separate deep neural network was trained to identify blood vessels, RBCs, and perivascular tissue region. These segmented “artifact” regions were subtracted from the prediction of the upper model during postprocessing to eliminate false positives. MNC indicates mononuclear cell infiltration; PCM, progressive cardiomyopathy; RBCs, red blood cells; WSI, whole slide image.

Figure 2.

Cardiomyopathy algorithm training and testing pipeline. The WSIs of heart sections from Sprague Dawley rat hearts were divided into tiles of size 512 × 512 pixels. The tiles from the heart WSI were used to generate a trained model by using the encoder–decoder network (see also Supplemental Figure 1). For testing, the WSI was again divided into tiles. Inference was performed on each of the tiles using the trained model. The tiles were then stitched back together to form the WSI with abnormality areas segmented and color coded. WSI indicates whole slide image.

Figure 3.

Scatterplot of AI scores by median grade. Violin plots (in gray shaded areas) are overlaid to show the shape of the AI distribution within each grade. A smooth (LOESS) curve (blue solid line) illustrates the nonlinear relationship between the scores predicted by the AI algorithm and the pathologists’ ratings. AI indicates artificial intelligence; LOESS, local polynomial regression.

Figure 4.

Scatterplot of log10-transformed AI scores at each median grade. Violin plots (in gray shaded areas) are overlaid to show the shape of the AI distribution within each grade. A smooth (LOESS) curve (blue solid line) illustrates the relationship between the log-transformed AI scores and the pathologists’ ratings. AI indicates artificial intelligence; LOESS, local polynomial regression.

Figure 5.

Sprague Dawley rat heart with color-coded segmentation areas. A-C, Algorithm output without the artifact pipeline. Fibrosis (red), mineralization (light blue), and MNC (yellow) were erroneously detected in the blood vessel. D-F, Algorithm output in the same sections with the artifact pipeline applied (after subtraction of blood vessel region, shown in blue). H&E stain, original scan 10×. H&E indicates hematoxylin and eosin; MNC, mononuclear cell infiltration.

Figure 6.

Sprague Dawley rat heart images before and after algorithm segmentation. A-C, Input images demonstrate microscopic features of PCM. D-F, The corresponding images following segmentation by the algorithm. Colored regions indicate areas predicted by the algorithm to have fibrosis (red), necrosis (green), mineralization (light blue), and MNC (yellow). Appropriate segmentation occurred even in areas of tissue fold artifact (E). H&E, original scan 5X. H&E indicates hematoxylin and eosin; MNC, mononuclear cell infiltration; PCM, progressive cardiomyopathy.

Figure 7.

Predicted probabilities of grade outcomes for given AI scores based on a multinomial logistic regression model of median grade as a function of the log10-transformed AI scores. AI indicates artificial intelligence; LOESS, local polynomial regression.