- Abstract: Object segmentation and classification in medical imaging are essential tasks for the diagnosis and understanding of diseases. Manual classification of the anatomical structures and the annotation of their boundaries are laborious tasks that also require strong medical expertise. Deep neural networks can be used to automate this task, but because of the unavailability of large datasets with multiple structures annotated and labeled, their performance is not at par with manual annotations. We propose a semi-automated interactive tool based on deep learning to produce high-quality annotations quickly. The architecture uses two convolutional networks: the first network produces multiple segmentations using a few clicks inside and outside the object, while the second classifies the object and selects one segmentation. We use MonuSAC histopathology data with four classes of labeled and annotated nuclei annotated as a testbed. On held-out images, our method was significantly more accurate in both segmentation and classification as compared to fully automated methods, while it was also at least 3 times faster as compared to manual annotation methods.
- Paper Type: well-validated application
- Track: short paper
- Keywords: Histopathology, Nucleus segmentation, Nucleus classication.
- TL;DR: An easy to use and fast semi-automatic tool to assist annotators with more accurate instance segmentation and classification in medical images.
- Supplementary Material: zip