Streamlining AI algorithms on remote HPC Infrastructure
Aim
Owing to the complexity and large size of image data in pathology, artificial intelligence (AI) algorithms preferably run on a dedicated high-performance cluster (HPC). Validation of newly developed algorithms usually requires integration into an image management system (IMS), which is a time-consuming process. We propose an IMS-independent approach to rapidly validate and deploy AI algorithms in research and diagnostic settings (Fig. 1). A generalizable interface for any image-related AI algorithm has been developed to set up an automatic workflow between the institute and the remote HPC. Once image data is generated, the data flow is managed and monitored by a data and job manager communicating with the HPC. A web-based graphical user interface (GUI) displays the status and allows pathologists to visually evaluate the result in a user-friendly manner and provide instant and structured feedback.
Figure 1: Integration workflow without image management system (IMS) [WSIs: whole slide images, HPC: high-performance computing, LIS: laboratory information system]
Methods
To validate the platform in a real diagnostic setting, we deployed it prospectively in the pathology clinic running MetAssist 1.0, a previously developed deep learning model that detects colorectal cancer lymph node metastases on H&E-stained whole slide images. Over the study period (May 2023 to September 2025), 1,783 lymph node slides from 139 patients were digitized, automatically routed to the HPC through the data and job manager, analyzed by the model, and returned to pathologists through the web interface, all alongside, and without modifying, the standard microscopy workflow. The model gave a positive or negative call per slide; whenever its prediction disagreed with the pathologist's initial reading, that slide was flagged for targeted re-review, turning the platform into a quality-control safety net.
To measure the effect of this assistance, we ran a reader study on a randomly selected sub-cohort of 256 slides with three readers, comparing three settings: unassisted review, AI-assisted review, and a first-read scenario in which pathologists examined only the slides the model flagged as positive. We recorded slide-level sensitivity and specificity, inter-observer agreement (Cohen's kappa, κ), and review time, and collected structured free-text feedback throughout prospective use to characterize the model's failure modes.
Results
The platform ran stably in daily routine and was well received by the 20 pathologists and residents who used it. After targeted re-review of the flagged slides, the workflow reached high slide-level performance, while its feedback loop recovered metastases that had initially been missed under the microscope:
| Metric (after AI-assisted re-review) | Value |
|---|---|
| Sensitivity | 99.5% (203/204; 95% CI 97.2–100%) |
| Specificity | 80.2% (1,266/1,579; 95% CI 78.1–82.1%) |
| Agreement with pathologists | ~82% |
| Initially-missed metastases recovered | 4 cases (0.2%) |
| Median false-positive slides per case | 2 (IQR 0–3), against ~12 slides per dissection |
In the reader study, AI assistance nudged pathologists into closer agreement, and reviewing only flagged slides made the workflow dramatically faster:
| Reader-study setting | Effect |
|---|---|
| Inter-observer agreement, assisted | up to Δκ = +0.05 |
| Inter-observer agreement, first-read | up to Δκ = +0.11 (κ reaching 0.94 between two readers) |
| Total review time | ↓ ~60–95% (e.g. from 61 to 3 minutes) |
Structured feedback from pathologists pinpointed the model's main false-positive triggers — tissue folds, vessel walls, fibrotic tissue, non–lymph node tissue, and tumor deposits outside nodes — and this feedback directly shaped the successor model, MetAssist 2.0.
Together, these results show that the platform can carry a deep learning model into routine diagnostics as a slide-level safety net — improving inter-observer agreement, cutting review time, and closing the loop from pathologist feedback to model improvement — with multicentre and health-economic studies the natural next step.
Members
Amjad Khan
Inti Zlobec
Bastian Dislich
Stefan Reinhard
Publications
Khan, A., Zens, P., Zagrapan, B., Reinhard, S., Perren, A., Garcia-Baroja, J., Dislich, B. & Zlobec, I. Artificial intelligence for colorectal cancer lymph node metastasis detection: clinical integration and prospective validation of MetAssist 1.0 (2026). [under review]