Abstract
Presently, pathologists need to stain biopsy samples with standard and antibody-based immunocytochemistry (ICC) reagents for final diagnosis. Antibody reagents take hours to days to perform staining, along with requiring specialized equipment and technical skills. We have developed an AI-based virtual ICC platform that measures individual cell morphological features in whole slide images and labels the cells as immuno-positive or negative. The platform runs on the cloud in minutes, saving pathologists significant time and cost. For this purpose, cytopathology slides were obtained from N = 100 suspected cases of canine T-cell and B-cell lymph node lymphomas through Fine Needle Aspiration (FNA). Cytopathology slides were initially stained with the standard Wright-Giemsa (WG) and then re-stained with ICC reagents, anti-CD3 or anti-PAX5 antibodies, resulting in a pair of stained slides (WG-CD3 or WG-PAX5). Prior to AI training, cytopathology slides were digitally scanned, and the resulting images underwent a comprehensive pre-processing protocol to separate stains of interest for nuclei segmentation in WG and CD3 or PAX5. Following nuclei segmentation, the cell features from processed image pairs were translated into a structured tabular features format with immuno-positive and negative labeled classes. In total, the geometrical features of 8.48 million segmented cells (4.24 million pairs) were translated into a tabular format and paired based on the Euclidean cell matching algorithm. This approach facilitated the prediction of cell labels, achieving sensitivity and specificity of 0.98 and 0.97 (0.94 and 0.99), respectively for CD3 (PAX5). Additionally, the AI-based virtual ICC has demonstrated capabilities in cell counting, cell spatial distribution, cell segmentation, and classification. It offers a rapid, accurate, and precise evaluation of FNA samples and has the potential to help advance diagnostic cellular and molecular pathology capabilities.
Introduction: The Bottleneck in Biopsy Diagnosis
Diagnosing cancers like lymphoma from Fine Needle Aspiration (FNA) biopsies typically begins with standard stains like Wright-Giemsa (WG). However, definitive subtyping—such as distinguishing T-cell from B-cell lymphoma—requires Immunocytochemistry (ICC) using specific antibodies like CD3 or PAX5. Physical ICC staining is a significant bottleneck in clinical workflows; it is labor-intensive, requires specialized laboratory infrastructure, and can take days to produce results.
This research introduces an AI-based "virtual ICC" platform designed to bypass these constraints. By analyzing the geometric morphology of cells in standard WG-stained images, the system predicts immuno-positivity with high accuracy, offering a rapid and cost-effective alternative to physical staining.
Methodology: A Pipeline for Virtual Staining
The core of this approach lies in a robust pipeline that translates raw image data into structured features for machine learning.
Perfectly Paired Imaging
To train the model, the researchers developed a method to create perfectly paired "before and after" images. FNA slides were first stained with standard WG and scanned. These same slides were then destained and re-stained with ICC reagents (CD3 or PAX5) and scanned again. This ensured that the AI was learning from the exact same cells across different staining conditions.
Data Processing and Feature Engineering
The resulting gigapixel images were aligned and processed through a multi-step framework:
- Registration & Stain Separation: Images were aligned using the SIFT algorithm, and color channels were normalized to isolate nuclei (Azure B for WG; blue/brown for ICC).
- Instance Segmentation: Individual cell nuclei were segmented using Stardist, a neural network optimized for star-convex polygons.
- Feature Extraction: Geometric properties such as area, diameter, and solidity were extracted into a tabular format.
- Cell Matching: WG-stained cells were matched to their ICC counterparts using a Euclidean distance algorithm to establish ground-truth labels.
The study utilized a dataset of 100 canine lymphoma cases, resulting in over 8.48 million segmented cells (4.24 million matched pairs) for training and validation.
Results: High Accuracy and Interpretability
Rather than using a "black box" deep learning model, the researchers employed a Random Forest Classifier (RFC) on the extracted geometric features. This approach proved highly effective:
- CD3 (T-cell marker): Achieved 97.5% accuracy, with a sensitivity of 0.98 and specificity of 0.97.
- PAX5 (B-cell marker): Achieved 97.8% accuracy, with a sensitivity of 0.94 and specificity of 0.99.
A key insight from this work is that basic nuclear morphology—specifically equivalent diameter, area, and major axis length—serves as a reliable proxy for molecular markers. Furthermore, the system provides visual outputs that localize positive and negative cells, aligning with the traditional workflows pathologists already trust.
Conclusion and Future Outlook
The AI-based virtual ICC platform demonstrates the potential to revolutionize diagnostic pathology by providing results in minutes rather than days. While currently limited by species specificity (canine samples) and occasional segmentation challenges with out-of-focus cells, the framework is highly adaptable. Future work will focus on extending this technology to other species and, crucially, translating these protocols for human clinical use in oncology.
Summary updated