Related Experiment Video
Updated: Sep 29, 2026
Automated Quantification of Hematopoietic Cell – Stromal Cell Interactions in Histological Images of Undecalcified Bone
Published on: April 8, 2015
ShadoNet: A Cell Detection and Classification Framework for Ki-67 Pathology Images
Mahsa Ghasemi1, Fuyong Xing2, Toby C Cornish3,4
1Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN 47405, USA.
Motivation:
Accurate detection and classification of individual cells are essential for quantitative histopathology. In Ki-67-based tumor grading, the clinically relevant index is calculated from immunopositive and immunonegative tumor cells, while non-tumor cells should be excluded. This is challenging because Ki-67-stained tumor regions often contain a heterogeneous mixture of immunopositive tumor cells, immunonegative tumor cells, and non-tumor cells. Here, we study cell detection and classification in Ki-67-stained histopathology images to support grading of pancreatic neuroendocrine tumors (PanNETs), where reliable tumor-cell-specific identification is critical for tumor grading and decision-making.
Results:
We present ShadoNet, a U-Net-style encoder-decoder framework that integrates shape priors with structured regression for cell detection and classification. ShadoNet predicts class-specific proximity maps that emphasize cell centers and jointly encode spatial location, morphology, and class identity through smooth intensity decay from cell centers. Across multiple Ki-67-stained histopathology datasets, incorporating shape information improved performance on standard detection and classification metrics, indicating that shape-derived priors provide useful class-related cues-without requiring full instance-level boundary annotations.
Availability And Implementation:
ShadoNet uses the Segment Anything Model (SAM) to generate initial cellular shape masks, which are refined using human-annotated nuclear point labels to construct ground-truth proximity maps. During training, shape guidance is enforced via two auxiliary objectives: (i) a rotation-extended Scaled IoU (SIoU) loss to capture orientation and shape alignment, and (ii) a Hausdorff Distance Transform (HDT) loss to penalize boundary discrepancies between predictions and ground truth. Code is publicly available at: https://github.com/GhasemiGOF/ShadoNet. A fixed version of the code used for the experiments in this manuscript has been archived at Zenodo with DOI: https://doi.org/10.5281/zenodo.21050210. BCData and SHIDC-B-Ki-67 are publicly available datasets and can be accessed through their original publications.
Supplementary Information:
Supplementary information is available for this manuscript.
