AI supports slide quality assessment: The Artifact Segmentation Model

“The aim of the project is to develop an AI model that detects artifacts using a semantic segmentation network,” says Radboudumc researcher Marina D’Amato, who is working on this topic as part of her PhD. “It’s designed to support quality control in large-scale pathology datasets by automatically flagging common image artifacts.”

What is the Artifact Segmentation Model?

At its core, this model is a deep learning-based semantic segmentation network that identifies and classifies six of the most frequent artifact types found in histopathology images: tissue folds, dust particles, out-of-focus regions, ink marks, pen markers, and air bubbles.

The model uses a two-step inference approach:

• Tissue segmentation: First, a segmentation network isolates tissue regions from the background to reduce computational load and focus on diagnostically relevant areas.
• Artifact segmentation: Next, the model identifies artifacts within these tissue regions. It outputs a mask that visually highlights artifact zones and classifies each pixel into one of the predefined categories.

Technically, the model architecture is built on DeepLab V3+ with an EfficientNet-B2encoder, chosen for its balance of performance and computational efficiency. This architecture is well-suited for large-scale slide processing, where memory constraints and runtime matter.

“This isn’t a novel architecture,” Marina explains, “but it’s fast, accurate, and efficient—key requirements for integrating this kind of tool into real-world workflows.”

The Artefact Segmentation Model in action (video)

A visual walkthrough of the Artifact Segmentation Model. Using ASAP, a whole slide image is overlaid with the predicted segmentation mask, allowing inspection of different artifact types and their locations.

Why is this model important for Bigpicture?

Artifacts in pathology slides are common and can arise during multiple stages (from mounting and staining to scanning) and vary depending on lab protocols and equipment. These artifacts compromise not only clinical diagnosis but also AI model training, as poor-quality inputs reduce accuracy and generalizability. “This model gives users an automatic quality control step. It flags low-quality slides, so users can take appropriate action or understand limitations,” says Marina.

Unlike other models trained on limited data (often only H&E stained slides), this model was trained with data diversity in mind. The initial version used 121 annotated slides representing 13 human tissue types, 9 stain types, and7 different scanner models.

The evaluation of the model involved over 500 cases, including preclinical, toxicological, and multi-species data, collected from Bigpicture partners such as Janssen, Novartis, and UCB. “The diversity of training data is what sets our model apart,” Marina says. “It generalizes well to unseen cases, which is crucial for deployment on a platform like Bigpicture.”

What can Bigpicture contributors expect from the AI model?

The Artifact Segmentation Model is designed to be integrated into the Bigpicture platform so that every newly uploaded slide can be scanned for artifacts. Concrete this would mean that the tool will flag poor-quality slides, assigns a quantitative quality score (e.g., 8.7% artifact coverage), and visualizes affected regions with an overlaid segmentation mask.

This doesn’t mean low-quality slides will be removed, as they may still be useful depending on context, but they’ll be clearly labeled. Marina: “We want to offer a realistic picture of data quality, not to filter out everything that’s imperfect. But users need to know what they’re working with.”

Ongoing work and future plans

The first version of the model performs well overall but shows weaker performance on certain artifact types, particularly air bubbles and mild blurriness. Retraining is underway using new data from Bigpicture partners to address these gaps. Marina: “We’re continuing to improve it by expanding and refining annotations and adding more diverse data.”

The model could potentially also support clinical workflows in labs. For example, in hospital settings where only 5% of slides are manually reviewed, this model could automatically assess all slides post-scanning, and recommend actions like “rescan” or “clean slide.”

Beyond quality control, the model also benefits downstream AI pipelines by excluding artifact-heavy regions.