Improving deep learning-based segmentation of diatoms in gigapixel-sized virtual slides by object-based tile...

Area of Science:

• Microbiology
• Eukaryotic cell biology
• Aquatic ecology

Background:

• Diatoms are diverse microscopic eukaryotes crucial for aquatic ecology and biomonitoring.
• Accurate identification and enumeration of diatoms rely on analyzing their silica shells (frustules).
• Digital analysis of diatom images requires robust segmentation, especially for large, high-resolution scans.

Purpose of the Study:

• To develop and evaluate deep learning-based segmentation methods for diatoms in gigapixel images.
• To investigate the impact of different tiling strategies on segmentation performance.
• To improve the discrimination of intact diatom frustules from fragments in complex backgrounds.

Main Methods:

• Applied deep learning models (Mask R-CNN, U-Net) to gigapixel diatom slide scans.
• Compared standard sliding window tiling with novel object-based tiling approaches (with and without integrity constraints).
• Trained models with varying amounts of data and analyzed segmentation performance metrics (precision, recall).

Main Results:

• Object-based tiling with an object integrity constraint significantly improved pixel-based precision by 12-17 percentage points.
• This improved precision was achieved without substantially compromising recall compared to standard tiling.
• The proposed tiling scheme enhances the ability to distinguish intact diatom frustules from irrelevant fragments.

Conclusions:

• Object-based tiling schemes, particularly with an integrity constraint, enhance diatom segmentation accuracy in large gigapixel images.
• This approach offers a more reliable method for automated diatom analysis in ecological and biomonitoring applications.
• The developed tiling strategies may also benefit segmentation tasks in other image analysis domains.