Vacuole Segmentation and Quantification in Liver Images of Wistar Rat

Accurate detection of macro and microvesicles in rat models of fatty liver disease is crucial in evaluating the progression of liver disease and identifying potential hepatotoxic findings during drug development. In this paper, we present a deep-learning-based framework for the segmentation of vacuoles in liver images of Wistar rat and study the correlation of automated quantification with expert pathologist's manual evaluation. To address the issue of misclassification of lumina (vascular and bile duct) as large vacuoles, we propose a selective tiling technique to generate tiles that include complete lumina and large vacuoles. A binary encoder-decoder convolution neural network is trained to detect individual vacuoles. We report a sensitivity of 85% and specificity of 98%. Furthermore, the diameter and roundness of the segmented vacuoles are estimated with an error of less than 8%, which supports the high potential of our method in drug development process.