[Paper] Aycromo: An Open-Source Platform for Automatic Chromosome Detection in Metaphase Images Based on Deep Learning

Source: arXiv - 2604.24685v1

Overview

The paper introduces Aycromo, an open‑source desktop application that brings state‑of‑the‑art deep‑learning chromosome detection to the hands of cytogeneticists. By wrapping pre‑trained models (e.g., YOLOv11) in a user‑friendly Electron interface, the tool turns a multi‑day, manual karyotyping process into a matter of seconds per slide—making AI‑assisted genetic diagnostics practical for everyday labs.

Key Contributions

• Fully open‑source platform (Electron + ONNX Runtime) that runs on Windows, macOS, and Linux without requiring any command‑line expertise.
• Plug‑and‑play model loading: users can drop any ONNX‑compatible detector (YOLOv8, YOLOv11, Faster‑RCNN, etc.) into the app and start inference instantly.
• Integrated benchmarking suite that automatically evaluates multiple models on a chosen dataset and visualizes mAP, inference time, and memory footprint.
• Interactive annotation & correction UI allowing experts to edit, add, or delete detections directly on the metaphase image, with changes saved back to a standard COCO‑style JSON.
• Performance validation on the CRCN‑NE metaphase dataset showing YOLOv11 reaches 99.40 % mAP@50 while processing a whole slide in under a second on a consumer‑grade GPU.

Methodology

1. Data preparation – The authors used the publicly available CRCN‑NE dataset, which contains high‑resolution metaphase spreads annotated with chromosome bounding boxes.
2. Model conversion – Popular PyTorch detectors were exported to the ONNX format, enabling hardware‑agnostic inference via ONNX Runtime.
3. Desktop integration – An Electron shell provides the UI; the heavy lifting (image loading, model inference, post‑processing) runs in a Node.js backend that calls ONNX Runtime.
4. Benchmark module – For each loaded model, the app runs a forward pass over a validation split, computes mean Average Precision at IoU = 0.5 (mAP@50), records latency, and presents results in sortable tables and charts.
5. Human‑in‑the‑loop correction – Detected boxes are rendered on a canvas where users can drag, resize, or delete them. The corrected annotations can be exported for downstream karyotyping pipelines.

Results & Findings

• Speed: Even the slowest model processes a 4 k × 4 k metaphase image in under 0.1 s on a mid‑range GPU, meaning a full slide (≈ 30 cells) can be analyzed in a few seconds.
• Usability: The UI eliminates the need for Python scripts or Docker containers; a cytogeneticist can start a new analysis with three clicks.
• Human correction: In pilot tests, experts spent < 5 seconds per slide to verify and adjust detections, a > 95 % reduction compared with manual karyotyping.

Practical Implications

• Clinical labs can integrate Aycromo into existing workflows to accelerate karyotype generation, freeing specialists to focus on interpretation rather than tedious box drawing.
• Research groups gain a ready‑made benchmarking environment to compare new detector architectures on chromosome data without writing boilerplate code.
• Software vendors have a reference implementation for building commercial AI‑assisted cytogenetics tools—thanks to the permissive MIT license and modular ONNX backend.
• Edge deployment: Because inference runs via ONNX Runtime, the same binaries can be shipped to low‑cost workstations or even high‑end laptops, removing the need for dedicated servers.
• Data standardization: Exported annotations follow COCO JSON, enabling seamless hand‑off to downstream karyotyping software (e.g., for chromosome banding classification).

Limitations & Future Work

• Dataset scope – Experiments are limited to the CRCN‑NE dataset; performance on other staining protocols (e.g., G‑band, Q‑band) or lower‑resolution images remains untested.
• Model diversity – Only object‑detectors were evaluated; the platform currently lacks support for segmentation‑based approaches that could capture chromosome contours more precisely.
• Hardware dependence – Real‑time speeds assume a modern GPU; CPU‑only inference is slower and may not meet the “seconds per slide” claim.
• Regulatory readiness – The tool is a research prototype; additional validation, audit trails, and compliance (e.g., FDA 21 CFR 820) are needed before clinical deployment.

Future directions outlined by the authors include adding a plug‑in system for chromosome classification models, extending the UI to support multi‑slide batch processing, and conducting multi‑center clinical trials to certify diagnostic accuracy.

Authors

• Jorge L. A. Lima
• Filipe R. Cordeiro

Paper Information

• arXiv ID: 2604.24685v1
• Categories: cs.CV
• Published: April 27, 2026
• PDF: Download PDF