Fig. 2

Examples of automated segmentation and occlusion sensitivity mapping. A Faster Recurrent Neural Network (FRCNN) was used for cell segmentation. First, it was trained by human example and after iterative learning, automated cell detection was performed (A). Segmented cells show a yellow elliptic border. With respect to explainable artificial intelligence, we used occlusion sensitivity mapping to retrace the decision-making process of the convolutional neural nets in image-level recognition (B). In occlusion sensitivity mapping, parts of the image are iteratively blocked from evaluation by the neural network and performance is measured. If the blocked part of the image is highly important for correct classifications, performance will drop accordingly. This process is iteratively repeated for the entire image. The result can be visualized in the sense that highly important image areas are highlighted (yellow/green) while less important or negligible areas are shaded (blue/purple)