27 Deep learning for automated image segmentation of in situ livers from organ retrieval photographs Free

Assessing hepatic steatosis during liver donation is crucial but challenging due to reliance on subjective visual assessments. Artificial intelligence (AI) models can offer an objective alternative, if computers can accurately identify the liver in the image. To date, AI models have depended on manual identification (image segmentation). We present the first automated segmentation model for in situ liver images.

Two independent anonymized image datasets were used. The first, with 309 images, was split 75:25 for training and testing. The second, with 71 images, was used for external validation. Ground truth annotations were verified by a transplant surgeon. We trained Detectron2 and YOLOv8, two open-source models, for automated segmentation.

Compared to human segmentations, in interval validation, Detectron2 and YOLOv8 achieved accuracies of 95% and 96%, and Areas Under the Receiver Operating Characteristic Curves (AUC) of 93% and 95%, respectively. In external validation, Detectron2 and YOLOv8 achieved accuracies of 98% and 97%, and AUC of 95% and 95%, respectively. Human annotation took a median of 43 (IQR:36-51) seconds. Detectron2 performed 48 times faster at 0.89 (0.83-0.96) seconds, while YOLOv8 was 469 times faster at 0.09 (0.09-0.10) seconds.

In situ segmentation is particularly challenging due to the liver’s similarity to surrounding tissues, unlike ex situ images with more distinct backgrounds. Deep learning models can achieve accurate and significantly faster segmentation than manual methods. Effective segmentation ensures that AI analysis remains focused on liver tissue, enhancing both the explainability and trustworthiness of AI-driven hepatic steatosis assessments.