Fig. 3.

Convolutional neural network (CNN) analysis and logistic regression model. (a) Analysis of a convolutional neural network trained to distinguish arteriolosclerotic ulcer of Martorell (ASUM) from controls, based on vessel-centred image crops. Grad-CAM (26) analysis for ASUM-prediction on an example image (upper left) shows the highest activation within the vessel wall (upper middle and right). Extraction of the 200 image areas of the highest and lowest ASUM-activation of all training and test images (lower left and middle) indicate more eosinophilic colour and hypocellularity as objective distinctive features. Manual review of highly ASUM-activated areas of the test images is shown on the right, (b) Receiver operating characteristic (ROC) curve of the logistic regression model with the diagnosis ASUM vs controls as the target variable and wall/lumen ratio, calcification, hyalinosis and cellularity as the explanatory variables (red) and with cellularity and calcification as explanatory variables (blue). (c) Evaluation of ASUM and controls using 2 criteria, cellularity index (Cel) <0.24 and presence of calcification (Cal). Purple box indicates patients, who fulfil at least 1 criterion. The blue box indicates patients, who fulfil none of the criteria. AUC: Area under the curve.