Source-linked AI summary
Nuclei instance segmentation and classification in histopathology images with StarDist
Martin Weigert, Uwe Schmidt
TL;DR
Nuclei instance segmentation and classification is challenging in densely packed histopathology images, motivating extension of STARDIST beyond fluorescence microscopy. The paper adds classification and related training and postprocessing strategies, evaluates them on Lizard and CoNIC data, and reports winning the preliminary and final CoNIC test phases.
Problem
The paper addresses nuclei instance segmentation and classification in histopathology, extending a method originally developed for fluorescence microscopy.
Method
The approach extends STARDIST with pixel-level class prediction and evaluates class balancing, color augmentation, test-time augmentation, ensembles, and shape refinement.
Results
The approach won the preliminary and final CoNIC test phases for the segmentation and classification task.
Takeaways & Limitations
STARDIST can be successfully used and extended for object instance segmentation and classification in histopathology.
Abstract
from arXiv · showhide
Instance segmentation and classification of nuclei is an important task in computational pathology. We show that StarDist, a deep learning nuclei segmentation method originally developed for fluorescence microscopy, can be extended and successfully applied to histopathology images. This is substantiated by conducting experiments on the Lizard dataset, and through entering the Colon Nuclei Identification and Counting (CoNIC) challenge 2022, where our approach achieved the first spot on the leaderboard for the segmentation and classification task for both the preliminary and final test phase.
1. INTRODUCTION
The paper investigates extending STARDIST from fluorescence microscopy to histopathology for nuclei instance segmentation and classification. It evaluates the extension on challenge data, finding that color augmentation, class balancing, test-time augmentation, ensembles, and shape refinement improve performance.
- STARDIST represents nuclei as star-convex polygons rather than bounding boxes, targeting densely packed, roundish objects.The paper examines whether this fluorescence-microscopy method can transfer to histopathology images.
- The extension adds nuclei classification to STARDIST and evaluates it on the Lizard dataset and CoNIC challenge data.The study also compares performance quantitatively with histopathology methods such as HoverNet.
- Color augmentations addressing hematoxylin and eosin staining variability are crucial for training models that generalize to new data.
- Class balancing is decisive because CoNIC metrics assign equal importance to each cell type.
- Test-time augmentations, model ensembles, and simple shape refinement substantially improve performance or segmentation quality.
2. METHOD
The method extends STARDIST with pixel-level class prediction, dataset balancing, augmentations, test-time aggregation, non-maximum suppression, shape refinement, and model ensembling. It is trained on imbalanced six-class Lizard patches and produces refined instance masks from star-convex polygon predictions.
- STARDIST detection and segmentation: A CNN predicts per-pixel object probabilities and radial distances, after which thresholded polygon votes are pruned by non-maximum suppression.The retained polygons represent individual object instances.
- Instance classification: A semantic segmentation head predicts per-pixel class probabilities, which are aggregated within each segmented object to assign its cell type.
- Data and class balancing: The Lizard training data contain 4981 256 × 256 × 3 images and six nucleus classes with severe class imbalance.Epithelial nuclei exceed 60% of objects, whereas neutrophil and eosinophil nuclei each comprise less than 1%.
- Data and class balancing: Simple oversampling is the most effective tested strategy for addressing the severe imbalance in nucleus types.The explored alternatives also include class-weighted loss and focal loss for the semantic segmentation head.
- Training and augmentation: Training uses geometric and pixel-wise color augmentations, including brightness, hue, and hematoxylin-eosin staining changes.The model uses a 64-ray STARDIST configuration with a depth-4 U-Net backbone.
- Test-time augmentation: Eight geometric test-time augmentations are merged by averaging predictions, while radial-distance channels are permuted to match transformed polar directions.
- Postprocessing: Shape refinement rasterizes winner polygons and their suppressed overlaps, then uses majority voting to form each instance mask.This modifies standard STARDIST postprocessing after non-maximum suppression.
- Model ensembling: Predictions from separately trained STARDIST models can be merged alongside test-time augmentations to form ensembles.
3. RESULTS
The experiments evaluate CoNIC metrics, ablations, and challenge submissions for nuclei segmentation and classification. Oversampling, staining augmentation, test-time augmentation, shape refinement, and ensembles contributed to strong leaderboard performance.
- Evaluation metrics: Multi-class panoptic quality mPQ averages panoptic quality across all T cell classes/types.Panoptic quality is the product of detection quality DQ and segmentation quality SQ.
- Ablation experiments: Simple oversampling yielded by far the best results for all metrics in the internal validation ablations.The baseline used oversampling, brightness + H&E staining augmentations, and no test-time strategy.
- Ablation experiments: Brightness augmentations were more effective than hue- or H&E-affecting augmentations on the internal validation data.The authors relate this observation to the strong similarity between internal training and validation data.
- CoNIC submissions: Staining augmentations considerably improved challenge performance, while test-time augmentations and shape refinement produced smaller additional gains.The differing augmentation results were observed between internal validation and external preliminary test data.
- CoNIC submissions: The three-model B–D ensemble achieved first place on the concluding preliminary leaderboard, and the four-model C–F ensemble took the top spot on the final leaderboard.Table 2 reports the preliminary and final leaderboard results for the EPFL | StarDist submissions.
4. DISCUSSION
The paper concludes that StarDist can be successfully extended for object instance segmentation and classification in histopathology. The approach won both the preliminary and final CoNIC challenge phases.
- StarDist was successfully used and extended for object instance segmentation and classification in histopathology.
- The approach won the preliminary and final test phases of the CoNIC challenge for segmentation and classification.
5. COMPLIANCE WITH ETHICAL STANDARDS
The study reports that no ethical approval was required.
- No ethical approval was required for this study.