A deep learning framework for quantification of hepatic steatosis in whole-slide images by tile-level scoring and screening

Objective. To develop a deep learning framework for the efficient and accurate quantification of hepatic steatosis in whole-slide images (WSIs). Approach. Two retrospective histopathological WSI datasets were collected and graded. A deep learning framework was proposed by incorporating tile scoring and screening, fat droplet segmentation, and WSI-level fat fraction (FF) calculation. Specifically, tile score was calculated based on morphological and color features, and used to screen tiles for subsequent segmentation. Six segmentation models were trained and compared, and the model with the highest Dice coefficient was used to automatically segment fat droplets within tiles. Tile-level FFs were yielded and aggregated through weighted averaging to obtain WSI-level FFs. Performances of the tile scoring and screening strategy, segmentation models and the tile matrix size were evaluated, and agreement with expert-assessed steatosis grades. Main results. Optimal tile scoring parameters were determined as a weight of 0.54 and a decision threshold of 0.23, achieving an F1-score of 0.97 and an area under the curve of 0.94, effectively distinguishing tiles with and without fat droplets. UNet++ exhibited the best segmentation performance (Dice = 0.97). With an optimal tile matrix size of 2048 × 2048 pixels, the integration of the scoring and screening strategy improved processing efficiency 5.6-fold, reducing the mean WSI processing time from 141.5 s to 25.2 s, while WSI-level FFs agreed with expert-assessed steatosis grades with a mean absolute error of 0.44%. Significance. The proposed framework enables accurate and efficient hepatic steatosis quantification, demonstrating high concordance with expert assessment and offering a promising clinical solution.