This is an old revision of the document!
🦠 ForamDeepSlice: AI-Powered Foraminifera Classification
Automated Fossil Species Identification from Micro-CT Scans Using Deep Learning
KVL Staff on Project
Abdelghafour Halimi
abdelghafour.halimi@kaust.edu.sa
Didier Barradas-Bautista
didier.barradas@kaust.edu.sa
Ronell Sicat
ronell.sicat@kaust.edu.sa
Building 1, Level 0, Office 0125
Collaborators
Ali Alibrahim
Physical Sciences and Engineering Division
ali.alibrahim@kaust.edu.sa
Abdulkader M. Afifi
Physical Sciences and Engineering Division
abdulkader.alafifi@kaust.edu.sa
Revolutionizing micropaleontology with AI!
Foraminifera are microscopic organisms whose fossilized shells reveal Earth's climate history. ForamDeepSlice uses deep learning to automatically classify 12 foraminifera species from micro-CT scans with unprecedented accuracy.
🎯 Key Results:
- 95.6% accuracy across 12 species
- 99.6% top-3 accuracy
- 109,617 2D slices from 97 specimens
- Interactive dashboard for real-time classification
📊 Research Overview
Dataset
- 97 micro-CT scanned specimens
- 27 species total, 12 selected for training
- 109,617 high-quality 2D slices
- Rigorous specimen-level data splitting
Methods
- 7 state-of-the-art CNN architectures tested
- Novel PatchEnsemble strategy
- Transfer learning from ImageNet
- Advanced data augmentation
Top Performing Models
| Model | Accuracy | F1-Score |
|---|---|---|
| ForamDeepSlice (Ensemble) | 95.6% | 95.0% |
| ConvNeXt-Large | 95.1% | 94.1% |
| NASNet | 93.7% | 92.5% |
| ResNet101V2 (Baseline) | 84.3% | 80.8% |
Most species achieved F1-scores exceeding 90%, with the best reaching 99.7%. Our PatchEnsemble approach significantly improved classification of challenging species.
💻 Interactive Dashboard
<color #000000> Classification Features:
- Drag-and-drop image upload
- Automatic preprocessing & segmentation
- Real-time species prediction
- Confidence scores & rankings
- Works with micro-CT and optical images
</color>
<color #000000> 3D Slice Matching:
- Find slice orientation in 3D models
- Advanced similarity metrics (SSIM, NCC, Dice)
- Interactive 3D visualization
- No coding required
</color>
🌟 Impact & Applications
| Scientific Contributions | Future Applications |
|---|---|
| First comprehensive micro-CT foram classification dataset | Expand to additional species |
| Accelerates fossil identification workflows (reduces expert time) | Mobile applications for field use |
| Supports biostratigraphy and climate reconstruction research | Integration with geological context |
| Fully reproducible Docker environment | Address homeomorphy challenges |
| Open-source framework for domain scientists | Integration with field equipment |
Note: ForamDeepSlice is a decision-support tool for experts, complementing—not replacing—careful taxonomic analysis.
📚 Resources
Acknowledgments: KAUST Core Labs, Supercomputing Core Lab (Ibex), Domingo Lattanzi-Sanchez (Geo-Energy Platform)