Fig. 4.

NuCLS model architecture. (a) The Mask R-CNN architecture was adapted for nucleus detection and classification, allowing some independence of the classification and detection tasks, which improves performance. (b) Other adaptations we made include: (i) supporting variable-size images at inference while preserving scale and aspect ratio; (ii) supporting hybrid training data that mixes bounding boxes and segmentations; (iii) simplifying object detection and (iv) generating full class probability vectors for each nucleus at inference