TL;DR
This paper introduces a modular deep learning workflow for classifying lymphoma subtypes based on the size of neoplastic nuclei in histological images. The approach combines advanced image processing with segmentation models, achieving high accuracy (up to 92%) and outperforming human pathologists in reproducibility and accuracy, while also demonstrating transferability to feline lymphoma.
What Problem Does This Paper Address?
Histopathological examination is crucial for diagnosing and subtyping lymphomas, with nuclear size being a key differentiator. However, manual assessment of nuclear size is subjective, time-consuming, and prone to inter- and intra-rater variability. Factors like varying nuclear shapes within a tumor and the absence of consistent size references (like red blood cells, which vary by species) further complicate accurate classification. This paper aims to address these challenges by developing an automated, reproducible, and interpretable workflow for classifying lymphoma based on nuclear size.
Methods
The core of the proposed solution is a modular workflow that separates image segmentation from classification.
- Nuclei Segmentation: The workflow utilizes deep learning models for segmenting individual nuclei. Two approaches were tested:
- Unet++: A semantic segmentation model trained specifically on canine lymphoma data.
- Stardist: A pre-trained, unmodified instance segmentation model, demonstrating the workflow's flexibility with different segmentation architectures.
- Feature Extraction and Filtering: After segmentation, nuclei are identified, and invalid objects (e.g., due to poor shape or connectivity) are filtered out using a "circularity" measure derived from scikit-image.
- Reference Size Determination: A novel reference size was established using the mean diameter of small non-neoplastic lymphocytes from canine and feline lymph nodes, considered more consistent than red blood cells.
- Classification Parameter Optimization: A brute-force parameter search was employed to find optimal thresholds for classifying nuclei into "small," "intermediate," and "large" size categories based on their diameter relative to the reference size.
- Final Image Classification: The overall image classification is determined by a majority vote of the classified nuclei within the image.
The modularity allows for interchangeable segmentation models and provides insights into each step, fostering user trust.
Datasets
- Canine Lymphoma Dataset: 116 cases were used, with 2-3 µm HE-stained sections digitized.
- A subset of 27 images with manually labeled nuclei was used for training and testing the Unet++ segmentation model.
- A separate dataset of 89 canine lymphoma images was used for parameter optimization and workflow testing. These images were labeled by three pathologists for image-level classification and also had individual nuclei annotated.
- Feline Lymphoma Dataset: 38 cases were used, with similar processing as the canine data. These were exclusively used for testing the transferability of the workflow.
- Reference Data: Small non-neoplastic lymphocytes from canine and feline lymph nodes were annotated to establish a reference size.
No new benchmark datasets were introduced; instead, existing histological samples were curated and annotated.
Key Insights and Ideas
- Modularity is Key: The workflow's strength lies in its modular design, separating segmentation from classification. This allows for flexibility in choosing segmentation models (Unet++, Stardist) and provides transparency into the process, enhancing user trust.
- Beyond Black Boxes: Even when using "black-box" deep learning models for segmentation, the workflow's post-processing steps (filtering, size comparison to a novel reference) offer interpretable insights into the classification decisions.
- Novel Reference Size: Using non-neoplastic lymphocytes as a reference size proved more robust than red blood cells, especially across species, addressing a significant limitation in previous methods.
- Transferability is Achievable: The workflow, optimized for canine data, demonstrated good transferability to feline lymphoma without specific feline model training, suggesting its broad applicability.
- Automated Classification Outperforms Manual: The automated workflow achieved higher accuracy and reproducibility compared to individual pathologists' assessments, highlighting the potential of AI in diagnostic pathology.
Key Results
- Canine Lymphoma Classification: The workflow achieved up to 92% accuracy on the canine test dataset, significantly outperforming the average accuracy of three individual pathologists (74.22%).
- Reproducibility: The automated approach provides reproducible results, unlike manual estimations which showed variability (ranging from 66.67% to 84% in accuracy for individual pathologists).
- Feline Lymphoma Transferability: The workflow achieved 81.57% accuracy with Unet++ and 84.21% accuracy with Stardist on feline lymphoma data, demonstrating its generalization capabilities.
- Segmentation Performance:
- Unet++ achieved a Dice score of 0.8379 and an F1 score of 0.8146 for object detection.
- Stardist achieved a Dice score of 0.7063 but a higher F1 score of 0.7993 for object detection, with a lower rate of invalid objects excluded (5.76% vs. 23.11% for Unet++).
Limitations and Future Work
- Manual Region Selection: The current workflow still relies on pathologists to manually select regions of interest within whole slide images, limiting full automation. Future work could focus on automated tumor region selection.
- Species Specificity: While transferability to feline lymphoma was demonstrated, the authors note that the workflow's applicability to other species has not been confirmed and would require repeating the modeling and parameter optimization.
- Data Requirements: The need for expert labels and individually marked nuclei for training presents a limitation due to data availability.
- Computational Cost: While efficient (25 images/minute on a Titan RTX), processing entire whole slide images might not always be beneficial if nuclear size is uniform across tumor regions.
Future research could explore automated tumor region selection, further optimization for different species, and potentially end-to-end classification models if interpretability is less of a concern.
Summary updated