Abstract
Cell detection is an essential task in cell image analysis. Recent deep learning-based detection methods have achieved very promising results. In general, these methods require exhaustively annotating the cells in an entire image. If some of the cells are not annotated (imperfect annotation), the detection performance significantly degrades due to noisy labels. This often occurs in real collaborations with biologists and even in public data-sets. Our proposed method takes a pseudo labeling approach for cell detection from imperfect annotated data. A detection convolutional neural network (CNN) trained using such missing labeled data often produces over-detection. We treat partially labeled cells as positive samples and the detected positions except for the labeled cell as unlabeled samples. Then we select reliable pseudo labels from unlabeled data using recent machine learning techniques; positive-and-unlabeled (PU) learning and P-classification. Experiments using microscopy images for five different conditions demonstrate the effectiveness of the proposed method.
Introduction: The Challenge of Imperfect Annotation
In biomedical computer vision, cell detection serves as the critical foundation for downstream tasks like counting and tracking. While deep learning has revolutionized this field, standard models are notoriously "data-hungry," typically requiring exhaustive annotations of every cell in an image. In practice, however, biologists often only have the time or resources to label a fraction of the cells.
This leads to a "noisy label" problem. When a model is trained on sparsely annotated data, standard algorithms mistakenly treat the remaining unlabeled cells as "background." Consequently, the model becomes confused as similar-looking objects are labeled as both "cell" (foreground) and "not-cell" (background), resulting in a significant drop in detection accuracy. This research aims to bridge the gap between easy-to-obtain sparse annotations and the high-performance requirements of biological research.
Proposed Methodology
The authors introduce an iterative five-step pipeline designed to "clean" training data and bootstrap model performance from imperfect labels:
- Initial Training with Masked Loss: A U-Net detection network ($D$) is trained using a "masked loss" function. This restricts learning to regions where annotations exist, preventing the model from being penalized for identifying unlabeled cells. While this avoids the background noise problem, it typically results in significant over-detection (false positives).
- Patch Generation: The model generates a heatmap for the entire image, and small patches are cropped around every detected point.
- Feature Extraction via PU-Learning: A second CNN ($\hat{g}$) is trained using Positive-and-Unlabeled (PU) learning. This technique is specifically suited for scenarios with confirmed positives and a mixture of unknown samples.
- Pseudo-label Selection via P-classification: To ensure high-quality labels, the framework employs P-classification. Unlike standard classifiers, P-classification is a ranking technique that prioritizes precision at the top of the list, ensuring that selected "pseudo-positives" are highly likely to be actual cells.
- Iterative Refinement: The most reliable pseudo-labels (top $\alpha\%$ for positives, bottom $\beta\%$ for negatives) are integrated into the training set, and the process repeats.
Experimental Setup and Datasets
The method was evaluated on two primary microscopy datasets:
- BF-C2DL-HSC (HSC): Hematopoietic stem cells from the Cell Tracking Challenge.
- C2C12: Myoblast cells cultured under four distinct conditions (Control, FGF2, BMP2, and a combination), presenting varied appearances that increase detection difficulty.
To simulate "imperfect annotation," the researchers provided labels for only approximately 10% of the total cells in the training images.
Results and Performance
The proposed method demonstrated substantial improvements over existing baselines for imperfect annotation. As shown in the results table below, the framework consistently outperformed other semi-supervised and masked-loss approaches.
Key findings include:
- Significant F-score Gains: On the HSC dataset, the baseline supervised method (Nishimura [12]) achieved an F-score of only 0.157 under missing label conditions. The proposed method boosted this to 0.952, nearing the 0.998 achieved with 100% exhaustive annotation.
- Robustness Across Conditions: Across all five conditions, the method maintained an average F-score of 0.780, nearly double the performance of the next best baseline (Yang [21] at 0.436).
- Efficiency: Performance gains typically saturated after just two iterations, indicating computational efficiency.
Conclusion and Future Work
By leveraging "over-detection" as a candidate pool and using ranking-based P-classification to avoid "poisoning" the training set with incorrect labels, this framework allows for high-accuracy cell detection with minimal annotation effort.
Future work may address the challenges of estimating the "prior" (expected cell proportion) in dense images for PU-learning and exploring dynamic adjustments for the selection thresholds ($\alpha$ and $\beta$). Additionally, testing the framework against extremely complex or artifact-heavy backgrounds remains an area for further investigation.
Summary updated