HoVer-NeXt
As a continuation of the CoNIC challenge, I developed a fully end-to-end pipeline for nuclei segmentation and classification of hematoxylin and eosin-stained whole slide images. This work tackled several important challenges, including optimizing the slow challenge code for high throughput, transferring the tile-based model to whole slide inference, and adapting the existing dataset with an additional class: mitoses. Find the full manuscript on Openreview. All figures are taken from the manuscript.
I included a self-training routine to include mitosis as an additional class into the lizard dataset, while also creating a separate new dataset that is automatically labelled by the trained model on lizard, but contains pHH3 ground truth annotations for mitosis.
The new pipeline was 5 times faster than the state-of-the-art, enabled to differentiation of mitoses and intraepithelial lymphocytes and compared to transformer-based approaches, does not suffer from tile normalization artefacts.
