Deep Parenchymal Steatosis Assessment and Quantification in Histopathology

Manual annotation of liver biopsies is time-consuming, prone to human error, and suffers from high inter-observer variability. Automated algorithms for classifying metabolic dysfunction-associated steatohepatitis (MASH) in liver transplants can improve accuracy, speed, and consistency in diagnosing hepatic steatosis. This study presents a hybrid approach that combines image processing, machine learning, and deep learning for steatosis classification, leveraging state-of-the-art segmentation models. The first step involves deep learning based segmentation models to obtain histological tissue detection. Next, Gabor kernels are applied, followed by an adaptive color deconvolution to separate hematoxylin and eosin channels, along with the Macenko method for extended steatosis detection. Then, the macrosteatosis quantification is done using various segmentation models. Microsteatosis cleaning is performed to reject false positives in microsteatosis with techniques such as logistic regression, one-class SVM, and Isolation Forest. The algorithm is developed and tested on 385 H&E-stained liver images obtained from 77 donors. It demonstrated improved performance in both tissue and steatosis detection, showing an overall accuracy of 0.54% higher and a remarkable 5.03% increase in F1 score as compared to the state-of-the-art HEPASS algorithm. The performance of the proposed model on another public dataset shows an improvement of over 3% in accuracy across all the sets.