Link to DOI – 10.1109/ISBI52829.2022.9761495.
2022, pp. 1-4
This work addresses cell instance segmentation in digital cytology slides. Automated segmentation of cells is an essential step toward automated cell analysis and pathology diagnosis, however cell instance segmentation remains a challenging task, especially for touching or overlapping cells. This work first introduces a new large dataset of overlapping urothelial cell Z-stacks, allowing the comparison and validation of modern algorithms for cell instance segmentation, improving previously proposed methods by a large margin. In addition, a modified backbone architecture is proposed to directly use Z-stacks as inputs, relieving from the necessity of using Extended Depth of Field projections of the volumes. Such backbone allows to reduce the total prediction time per sample by a factor up to 8, with no drop in segmentation performances. Dataset and code are available at gitlab.com/vitadx/articles/zstacks_cell_instance_segmentation.