TL;DR
This paper introduces ActiveSSF, a novel framework that combines active learning with self-supervised learning to accurately classify megakaryocyte subtypes from stained blood slides. By intelligently filtering noisy images, adaptively selecting informative samples, and leveraging prototype clustering, ActiveSSF significantly improves the recognition of rare cell subtypes, a critical step for diagnosing diseases like myelodysplastic syndrome.
What Problem Does This Paper Address?
Diagnosing myelodysplastic syndromes (MDS) relies heavily on the accurate classification of megakaryocytes, crucial cells in blood formation. However, analyzing these cells from stained slides presents several significant challenges for automated systems:
- Noisy Images: Blood smear images are often cluttered with background noise and non-uniform illumination, obscuring the fine details of the megakaryocytes.
- Long-Tailed Distribution: Like many real-world datasets, megakaryocyte subtypes are not evenly distributed. Some subtypes are very common, while others are extremely rare, making it difficult for models to learn effectively from limited data for these rare cases.
- Complex Morphology: Megakaryocytes exhibit a wide range of shapes and appearances within the same subtype (high intra-class variability), making it hard to distinguish them from each other and from non-target cells.
Existing deep learning and self-supervised learning methods struggle with these combined issues, particularly in achieving high accuracy for rare subtypes.
Methods
ActiveSSF tackles these challenges through a two-stage, active-learning-guided self-supervised pretraining framework:
- Stage 1: Cell Region Filtering:
- Noise Suppression: Images are first preprocessed with Gaussian filtering to reduce high-frequency noise.
- K-Means Clustering: A two-stage K-means clustering approach is used to segment the image into distinct color regions, effectively separating cellular material from the background.
- HSV Color Space Analysis: The filtered images are converted to the HSV color space, which is more effective for distinguishing cellular features. Clinical prior knowledge is used to define specific HSV thresholds for identifying potential megakaryocyte regions.
- Morphological Filtering: Candidate regions are further refined using morphological constraints (minimum bounding box size and fill rate) to eliminate false positives.
- Stage 2: Active Sample Selection:
- Prototype Generation: Labeled megakaryocyte samples are passed through a ResNet backbone to extract deep features. These features are then clustered using K-means to generate representative prototypes for each of the 11 subtypes.
- Adaptive Thresholding: A novel "density-aware" threshold mechanism is introduced. This mechanism dynamically adjusts the similarity threshold for selecting unlabeled samples. For common subtypes (large clusters), stricter thresholds are applied, while for rare subtypes (small clusters), more lenient thresholds are used. This ensures that informative samples from rare classes are more likely to be selected for pretraining.
- Self-Supervised Pretraining: The selected samples are then used for self-supervised pretraining, leveraging the learned representations to improve downstream classification performance.
Datasets
The study utilizes a comprehensive clinical dataset comprising:
- 7,577 labeled images of megakaryocytes, covering 11 distinct subtypes.
- 10,000 unlabeled blood cell slides for self-supervised pretraining.
The dataset is noted to have a pronounced long-tailed distribution, with "Granular megakaryocyte (G-MK)" being dominant and "Megakaryocyte with cytoplasmic abnormalities (CA-MK)" and "Large megakaryocyte with highly lobulated nucleus (HL-MK)" being rare.
Key Insights and Ideas
- Active Learning for Pretraining: The core innovation is using active learning principles during the self-supervised pretraining phase, not just for downstream fine-tuning. This actively guides the model to focus on the most informative unlabeled data.
- Adaptive Thresholding for Long-Tails: The dynamic threshold adjustment is a clever way to combat class imbalance. By lowering the bar for rare classes, the framework ensures they contribute meaningfully to the pretraining, preventing them from being overlooked.
- Clinical Priors Enhance Robustness: Integrating clinical expertise to define HSV thresholds for cell region detection demonstrates the value of domain knowledge in improving the robustness of image preprocessing.
- Prototype-Guided Selection: Using learned prototypes from labeled data as anchors for selecting unlabeled samples provides a principled way to ensure that the pretraining data is relevant and representative of the target classes.
Key Results
- ActiveSSF consistently improves performance across various state-of-the-art self-supervised learning methods (e.g., MAE, ConvMAE, MoCo v3, MIXMIM).
- When integrated with ConvMAE, ActiveSSF achieves the best overall performance.
- The framework shows significant improvements in classifying rare megakaryocyte subtypes, directly addressing the long-tailed distribution challenge.
- Visualizations (Figure 3) highlight substantial performance gains, particularly for rare subtypes, indicating ActiveSSF's effectiveness in mitigating class imbalance.
Limitations and Future Work
- Limitations: The paper acknowledges the inherent challenges of complex morphological variations and the long-tailed distribution, which ActiveSSF aims to mitigate but may not entirely eliminate. The reliance on specific HSV thresholds might require adaptation for different staining protocols or imaging conditions.
- Future Work: The authors suggest extending the ActiveSSF framework to other medical imaging domains that also suffer from scarce labeled data and imbalanced class distributions. Further exploration into adapting the framework for different staining techniques and imaging modalities is also a natural next step.
Summary updated