Computer Vision News - April 2022

19 Christian Marzahl human and feline samples. Afterwards, multiple clustering and manually screening steps were incorporated into the annotations to increase the overall dataset quality (Figure 1). My research showed that, for pulmonary haemorrhage detection, we were able to bridge the domain gap between equine and human samples [ 4 ] . Supporting pathologists with deep learning-based methods: Digitisation forms the basis for supporting the work of pathologists by utilising computer-assisted methods. Modern machine learning methods can accelerate previously time-consuming quantitative analyses, support reporting, and, therefore, standardise and improve the results of pathological examinations.Hence, I developednovel methods and software solutions that canbeusedtosupportboththeanalysis of image data and the cooperation between users. Figure 2 presents the results of the developed deep learning- based object detection method to quantify pulmonary haemorrhage on whole slide images by classifying macrophages into five corresponding grades, represented in the figure by bounding boxes with unique colours. Furthermore, my work indicated that experts are biased towards accepting precomputed annotations, which has important implications for using AI in clinical routines in the future [ 5 , 6 ] . Finally, to support reproducible scientific research, all code developed during my Ph.D. and the datasets used are publicly available on GitHub [ 7 ] . Figure 1. Visualisation of the developed pipeline to create high-quality multi-species datasets [ 4 ] Figure 2. AI object detection and annotations of pulmonary haemorrhage on whole slide images.