Deep learning and digital pathology powers prediction of HCC development in steatotic liver disease

Abstract

Background and aims: Identifying patients with steatotic liver disease who are at a high risk of developing HCC remains challenging. We present a deep learning (DL) model to predict HCC development using hematoxylin and eosin-stained whole-slide images of biopsy-proven steatotic liver disease.

Approach and results: We included 639 patients who did not develop HCC for ≥7 years after biopsy (non-HCC class) and 46 patients who developed HCC <7 years after biopsy (HCC class). Paired cases of the HCC and non-HCC classes matched by biopsy date and institution were used for training, and the remaining nonpaired cases were used for validation. The DL model was trained using deep convolutional neural networks with 28,000 image tiles cropped from whole-slide images of the paired cases, with an accuracy of 81.0% and an AUC of 0.80 for predicting HCC development. Validation using the nonpaired cases also demonstrated a good accuracy of 82.3% and an AUC of 0.84. These results were comparable to the predictive ability of logistic regression model using fibrosis stage. Notably, the DL model also detected the cases of HCC development in patients with mild fibrosis. The saliency maps generated by the DL model highlighted various pathological features associated with HCC development, including nuclear atypia, hepatocytes with a high nuclear-cytoplasmic ratio, immune cell infiltration, fibrosis, and a lack of large fat droplets.

Conclusions: The ability of the DL model to capture subtle pathological features beyond fibrosis suggests its potential for identifying early signs of hepatocarcinogenesis in patients with steatotic liver disease.

Graphical abstract

FIGURE 1

Flow diagram of the study population. Abbreviation: SLD, steatotic liver disease.

FIGURE 2

Schematic representation of the image-based deep learning model. (A) Method for model development using the paired case group. Patches were cut in order from top left to bottom right in a 256 × 256 size and divided into 5 groups. The cropped images from paired patients were included in the same group. Four of the 5 groups of cropped images were used as training and validation data sets, and the remaining group was used as the test set. After training, the classifier was evaluated using a test set. Five classifiers (CNN1–5) were created by repeating this process 5 times, using alternating training, validation, and testing sets. (B) Prediction of the probabilities of HCC development. The classifier based on deep convolutional neural networks (CNN) calculated the predictive cancerous value for the patches. The predicted probability of the HCC class is shown in each tile, where red represents ≥60%, gray represents between 40% and 60%, and blue represents <40%. The probability of each tile was averaged across the entire image to give the final probability. (C) Method for test with a model ensemble using the nonpaired case group. Each of the 5 classifiers calculated predictions for the test image in a similar way to (A). Those predictions were soft-ensembled into the final predictions of the image.

FIGURE 3

Predictive ability for HCC class. The confusion matrix and AUC for each group are shown for (A) predicted results of cross-validation using the image-based DL model on the paired case group; (B) predicted results of ensemble using the image-based DL model on the nonpaired case group as an additional test; (C) predicted results of cross-validation using the logistic regression model on the paired case group; and (D) predicted results of ensemble using the logistic regression model on the nonpaired case group as an additional test.

FIGURE 4

Saliency map. (A) Predicted HCC class cropped images (left images) and its saliency maps (right images). The darker the red color, the more salient the predicted tumorigenic feature is. (B) Predicted non-HCC class cropped images (left images) and its saliency maps (right images). The darker the red color, the more salient the predicted nontumorigenic feature is.