Artificial intelligence assisted compositional analyses of human aortic aneurysms ex vivo

Histological compositional information in relation to biomechanical parameters may provide essential information for developing models for human aortic aneurysm (AA) rupture prediction.
We hypothesize that artificial intelligence can support and improve the processing of histological information of excised human AA specimens. We present an analysis framework for automatic determination of tissue constituents within serial sections of excised AA specimens.
We used TensorFlow and QuPath to determine the overall architecture of the human AA: thrombus, arterial wall (primarily vascular smooth muscle) and adventitial loose connective tissue. Within the wall and adventitial zones, the content of collagen, elastin, and inflammatory (among others CD68+) cells was quantified. A deep neural network (DNN) based on the U-Net architecture, pretrained on the ImageNet dataset, was trained on manually annotated, Weigert stained AA tissue sections (17 patients). The DNN was able to segment the sections according to the overall architecture with Jaccard coefficients (after 65 epocs) of 77% for the training (annotated) and 76% for the validation (2 additional, new patient images) data, respectively. These segmentations were used in QuPath to facilitate quantification of per zone cell content and elastin/collagen area ratio. These latter measures were deterministically determined, and QuPath entries are provided for end users.
Our method can accurately outline the overall architecture of the human AA and recognize selected immune response cells and matrix constituents in each zone. We speculate that such data in combination with mechanical information can inform computational models predicting AA rupture.