2D U-NET Deep Learning Architecture for Atherosclerotic Lesion Segmentation and Quantification

Abstract

Mice are commonly used for experimental atherosclerosis research, and a common procedure involves morphometric analysis of atherosclerotic lesions on Oil Red O-stained cross-sections of the aortic root. However, manual measurement of atherosclerotic lesion sizes on numerous histological sections is tedious and time-consuming. Here, we developed a deep learning model based on convolutional neural network to quantify atherosclerotic lesions in cross-sectioned aortic roots of mice. Microscopy images were taken from cross-sectioned aortic roots of genetically diverse F2 mice derived from BALB/cJ and LP/J Apoe knockout mice and fed 12 weeks of a Western diet. Aortic cross-sections were stained with Oil Red O and hematoxylin and images captured. Atherosclerotic lesions were manually segmented in 2553 images using Zeiss AxioVision. These images were split into 1,533 for training the U-NET model, 510 for validation, and 510 images for testing. The developed model was evaluated for its performance to measure atherosclerotic lesions by comparing with manual measurement in test images. The model achieved a high degree of accuracy, with an F1 score of 0.82, a Jaccard score of 0.72, a Recall of 0.80, a precision of 0.87, and a correlation coefficient value of 0.98. While manual lesion measurement took approximately 5 minutes per image, the model performed the task in ~3.35 seconds, reducing measurement time by 98.9%. The developed 2D U-NET model provides a new platform for rapid and accurate quantification of atherosclerotic lesion sizes on Oil Red O-stained sections.