🤖 machine learning

KAYRA: A Microservice Architecture for AI-Assisted Karyotyping with Cloud and On-Premise Deployment

This paper introduces KAYRA, a containerized microservice architecture for AI-assisted karyotyping that combines a cascaded deep learning pipeline with flexible cloud or on-premise deployment, demonstrating statistically significant improvements in segmentation and classification accuracy over existing commercial systems in a pilot clinical evaluation.

Original authors: Attila Pintér, Javier Rico, Attila Répai, Jalal Al-Afandi, Adrienn Éva Borsy, András Kozma, Hajnalka Andrikovics, György Cserey

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Attila Pintér, Javier Rico, Attila Répai, Jalal Al-Afandi, Adrienn Éva Borsy, András Kozma, Hajnalka Andrikovics, György Cserey

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a hospital lab where scientists need to sort through a messy pile of tiny, tangled strings (chromosomes) to check for genetic health. This process, called karyotyping, is like trying to untangle a knot of Christmas lights while blindfolded, then sorting them by color and size. Traditionally, highly trained experts do this manually, taking 30 to 90 minutes per patient. It's slow, expensive, and hard to find enough experts to do it.

KAYRA is a new software system designed to act as a "super-assistant" for these experts. It uses artificial intelligence to do the heavy lifting, but it's built in a very specific way to fit into real hospitals.

Here is how KAYRA works, explained through simple analogies:

1. The "Swiss Army Knife" of Software (Microservices)

Most software is like a giant, heavy toolbox where if one tool breaks, the whole box is useless. KAYRA is different. It is built like a modular kitchen.

Instead of one giant robot doing everything, KAYRA is made of small, independent "chefs" (microservices) working in a line.
One chef cuts the vegetables (finds the chromosomes), another chops them (separates them), another sorts them by type, and a fourth arranges them on a plate.
Why this matters: If the "chopping" chef gets stuck, the "sorting" chef can still keep working with what they have. The system doesn't crash; it just gives the human expert a partial result to fix. This is crucial in a hospital where you can't afford for the machine to just stop working.

2. The "Zoom-In" Strategy (Cascaded ROI)

Looking at a whole microscope slide to find chromosomes is like trying to find a specific needle in a haystack by looking at the whole barn. It's too much information.

Step 1 (The Otsu Filter): KAYRA first takes a quick, blurry look at the whole image to find the "haystack" area. It ignores the empty space.
Step 2 (The U-Net Refiner): It then zooms in closer, using a smart filter to clean up the edges and define exactly where the chromosomes are.
Step 3 (The Mask R-CNN): Finally, it zooms in tight on just the chromosomes. Now, the AI doesn't have to guess; it only has to separate the tangled strings from each other.
The Benefit: By narrowing the focus step-by-step, the AI makes fewer mistakes because it isn't distracted by the background.

3. The "Double-Door" Deployment (Cloud vs. On-Premise)

Hospitals have strict rules about patient data. Some can send data to the "cloud" (the internet), but others cannot because the data must stay inside the building.

KAYRA is like a universal suitcase. It is packaged in a way that it can be unpacked and run in two different rooms:
- The Cloud Room: It runs on powerful, remote servers (like a hotel).
- The On-Premise Room: It runs on a local computer inside the hospital (like a home office).
The "suitcase" (the software code) is exactly the same in both rooms. This means hospitals that are strict about privacy can still use the same powerful AI as those that are not.

4. The "Human-in-the-Loop" Review

KAYRA doesn't replace the human expert; it acts like a draftsman.

The AI does the initial sorting and labeling.
Then, a human expert looks at the screen. They can see the AI's work and easily fix mistakes: "No, that's not chromosome 5, it's chromosome 10," or "These two are stuck together, please separate them."
The system remembers every change the human makes, creating a perfect history of how the final result was reached.

5. The Results: How Good is It?

The researchers tested KAYRA against two other systems (one old, one modern) using 459 chromosomes from 10 patient samples.

Separating the strings: KAYRA got it right 98.9% of the time. The old system got it right only 40% of the time, and the modern AI competitor got it right 78% of the time.
Sorting by type: KAYRA correctly identified the chromosome type 89.1% of the time. This was better than the old system (54.5%) and slightly better than the modern AI (86.9%), though the difference with the modern AI wasn't statistically huge yet.
Orientation: KAYRA got the rotation right 89.7% of the time. It was better than the old system but slightly behind the modern AI (94.5%).

The Bottom Line

The main point of this paper is that you can build a powerful, complex AI system for medicine that is flexible enough to work anywhere (cloud or local) and robust enough to handle mistakes without crashing. It proves that a multi-step AI pipeline can be packaged into a reliable service that helps doctors work faster and more accurately, while still keeping the human expert in charge of the final decision.

Note: The paper states this system is currently at "TRL 6" maturity, meaning it has been demonstrated in a relevant environment (a pilot study) but is still being prepared for full regulatory approval (like FDA or CE marking) for widespread clinical use.

1. Problem Statement

Karyotyping is a critical clinical diagnostic procedure involving the visual identification, ordering, and pairing of human chromosomes from G-banded metaphase microscope images. It is essential for prenatal screening, oncology (e.g., detecting the Philadelphia translocation), and hereditary disease diagnosis.

Current Challenges: The process is currently manual, time-consuming (30–90 minutes per case), and relies on a scarce workforce of highly trained cytogeneticists.
System Engineering Gap: While deep learning models (e.g., Mask R-CNN, U-Net) have shown high accuracy in segmentation and classification, there is a lack of robust system engineering solutions. Specifically, there is a need for a deployable clinical service that:
- Handles multi-model pipelines (detection, segmentation, classification).
- Operates within strict clinical constraints (patient data residency, on-premise vs. cloud flexibility).
- Integrates a "human-in-the-loop" workflow for expert review.

2. Methodology: The KAYRA System

KAYRA is an end-to-end karyotyping system architected as a containerized microservice pipeline. It is designed to be deployment-agnostic, running identical container images in both cloud (managed Kubernetes) and on-premise (Docker Compose) environments.

A. System Architecture

The system decomposes the workflow into single-responsibility microservices communicating via HTTP/REST:

Pipeline Orchestrator: Coordinates the flow per image. It implements degraded operation: if an upstream service fails, the pipeline skips that stage and returns partial results rather than failing entirely, ensuring utility for the cytogeneticist.
Model Services:
1. Semantic Segmentation: EfficientNet-B5 + U-Net.
2. Instance Detection: Mask R-CNN (ResNet-50 + FPN).
3. Duplicate/Overlap Resolution: A dedicated logic stage.
4. Classification + Rotation: ResNet-18 classifier.
Backend: Handles image ingestion, asynchronous task queuing, interactive editing APIs, and ISCN-compliant karyogram synthesis.
Frontend: Supports interactive expert review (splitting, merging, reclassifying, rotating chromosomes) with full audit trails.

B. The Cascaded ROI-Narrowing Pipeline

A core innovation is the cascaded Region of Interest (ROI) strategy, which applies a "divide-and-conquer" approach to handle overlapping chromosomes and reduce computational load:

Otsu Pre-filter: The original image (e.g., 1830×1830) is downsampled, binarized using Otsu's method, and connected components are labeled to create a coarse crop (crop1).
Semantic Refinement: crop1 is downscaled to 992×992 and processed by a U-Net (EfficientNet-B5 backbone) to generate a 3-class mask (background, chromosome, overlap).
Tight ROI Extraction: The non-zero region of the U-Net mask defines a tighter crop (crop2) at original resolution.
Instance Segmentation: Mask R-CNN runs only on crop2 (at 0° and 45° angles) to detect individual chromosomes and generate pixel-precise masks.
Classification: Detected chromosomes are cropped and passed to a ResNet-18 classifier for chromosome type (1–22, X, Y) and rotation angle estimation.

Why the Cascade Matters:

Efficiency: Mask R-CNN processes a ~34% smaller pixel area, reducing RPN proposal load.
Stability: Fixed input sizes (992×992) stabilize memory profiles and batching.
Accuracy: Each model sees only the data it is best suited to process (e.g., U-Net removes background noise before Mask R-CNN attempts to separate overlaps).

C. Training Data

Models were trained on a combined dataset of ~24,000 expert-annotated karyograms, including:

Peripheral blood and bone marrow samples.
Structurally abnormal cases (translocations, deletions).
Specific aberrations (e.g., t(9;22) Philadelphia chromosome).
Data from commercial platforms for cross-vendor robustness.
Privacy: Data partitioning is done at the patient-identifier level to prevent leakage between train/validation/test sets.

3. Key Contributions

Microservice Architecture: A flexible, containerized system supporting both cloud and on-premise deployment, addressing the critical requirement of patient data residency.
Cascaded ROI Pipeline: A novel engineering approach that chains Otsu pre-cropping, U-Net semantic refinement, and Mask R-CNN instance segmentation to improve performance on overlapping chromosomes.
Clinical Validation: A rigorous pilot study comparing KAYRA against two commercial reference systems (one density-thresholding, one modern AI) on real clinical data.
Human-in-the-Loop Integration: A workflow designed for clinical adoption, allowing cytogeneticists to correct AI outputs with full versioning and auditing.

4. Results

The system was evaluated on 459 chromosomes from 10 metaphase spreads (7 normal, 3 abnormal) against two commercial references.

Metric	KAYRA	Ref 2 (Modern AI)	Ref 1 (Density Threshold)	Statistical Significance (KAYRA vs. Ref 1)
Segmentation Accuracy	98.91%	78.21%	40.52%	$p < 0.0001$
Classification Accuracy	89.1%	86.9%	54.5%	$p < 0.0001$
Rotation Accuracy	89.76%	94.55%	78.43%	$p < 0.0001$

Segmentation: KAYRA significantly outperformed both references, correctly segmenting 454/459 chromosomes. It notably reduced "merged" errors compared to Reference 2.
Classification: KAYRA achieved 89.1% accuracy, significantly beating the older Reference 1. The difference vs. Reference 2 (86.9%) was not statistically significant ( $p=0.34$ ) given the sample size, though KAYRA showed superior performance on medium-sized B and C groups.
Rotation: KAYRA lagged slightly behind Reference 2 (89.76% vs. 94.55%), identified as an area for improvement via training data augmentation.
Robustness: KAYRA performed better on technically difficult bone-marrow samples (93.5% accuracy) than on peripheral blood (86.5%), likely due to the training distribution.

5. Significance and Conclusion

Clinical Readiness: KAYRA has reached TRL 6 (Technology Readiness Level 6), demonstrating a validated prototype in a relevant environment.
Deployment Flexibility: The architecture proves that complex multi-model AI pipelines can be packaged to satisfy strict clinical data governance (on-premise) without sacrificing the scalability of cloud deployment.
Future Work: The system is backbone-agnostic, allowing future upgrades to transformer-based models (e.g., Mask2Former) without architectural overhaul. Planned next steps include multi-centre validation, improving rotation accuracy, and addressing the under-representation of the Y chromosome.

Conclusion: KAYRA successfully bridges the gap between academic deep learning models and deployable clinical cytogenetic services, offering a robust, flexible, and highly accurate solution that integrates seamlessly into existing expert workflows.