Artificial intelligence assisted compositional analyses of human abdominal aortic aneurysms ex vivo

FIGURE 1. Data analysis flowchart. (A) The Deep Neural network is pretrained on ImageNet and trained on basis of manually annotated images of Weigert stained tissue sections of excised AAA specimens from 14 patients. The manual annotation defines the ground truth for the network. The ability of the network to recognize the different defined zones of the AAA wall is validated against manually annotated images from two other patients (B) In parallel, QuPath is used for detection of cells and elastin and collagen fibers (right). Finally, by overlaying the defined AAA wall zones and the QuPath output, quantification of the content of specific cells and extracellular matrix fibers per zone is possible.

FIGURE 2. Example of an input tile and a target tile from the mosaic creation process. (A) input tile, (B) target tile. Both tiles are 384 × 384 pixels (∼450 nm/pixel), which gives an adequate amount of information about the tissue for determining the class label(s) in the field of view.

FIGURE 3. Example of raw, annotated and segmentation prediction image. (A) Input image, Weigert stained (B) Partially annotated ground truth. The gray area surrounding the purple region is removed from the dataset during training to reduce class imbalance and improve training time (C) Segmentation prediction output from the neural model. Color coding: Blue: Zone 1. Green: Zone 2. Red: Thrombus. Purple: Background. Gray: Ignore.

FIGURE 4. The basic building block of our U-Net variant and the overall network architecture. (A) The BSConv2 block consists of two repeats of a reflection padded 3 × 3 convolution and linear activation followed by batch normalization and a trainable Swish activation (B) The U-Net architecture is structured into levels that share operation parameters. For a U-net with a depth of three, there are three levels below the surface level (depth 0). For each subsequent level, the number of filters per convolution is doubled and the spatial dimensions are halved.

FIGURE 5. Training and validation metrics per epoch. The Jaccard similarity index (JSI) calculated during training and validation is reduced because of a misconfiguration (see Discussion for details). The F1 and Jaccard scores for training are overall mean measures of the models’ precision and recall. Loss is a measure of costs of the model, and represents how poor the model is.

FIGURE 6. Input and output comparison. (A) Input images, Weigert stained (B) Segmentation prediction output from the neural model. Color coding: Blue: Zone 1. Green: Zone 2. Red: Thrombus. Purple: Background. Gray: Ignore.

FIGURE 7. Confusion matrix of the performance of the network model on the validation dataset. The numbers denote the fraction of pixels classified as label x and belonging to label y. The matrix is normalized by true label, i.e., each row summates to 1. The diagonal elements represent true positives, while off-diagonal elements represent false positives (columns) and false negatives (rows).

FIGURE 8. Total area and elastin and collagen area percentages in zones 1 and 2. (A) zone area for zone 1 (vascular wall) and zone 2 (perivascular loose connective tissue) (B) area percentages for elastin and collagen in zone 1 and 2, respectively.

FIGURE 9. Illustration of cell detection using antiCD8 immunostaining. (A) Raw immunostained image (DAB and hematoxylin stained). (B) Image in A with overlaid zone predictions (Blue: Zone 1 border. Green: Zone 2 border. Red: Thrombus) and all cell detections (Light blue: Hematoxylin stained nucleus in CD8 negative cell. Light red: Hematoxylin-stained nucleus in CD8 positive cell). (C) zoom in of square in (B). Scalebar is 1 mm (A, B) 250 µm (C) Arrow: part of the tissue that we can assume would be zone 1. Please refer to the discussion for details.