ForamDeepSlice: A High-Accuracy Deep Learning Framework for Foraminifera Species Classification from 2D Micro-CT Slices

This study introduces ForamDeepSlice, a high-accuracy deep learning framework that combines an ensemble of ConvNeXt-Large and EfficientNetV2-Small models with a rigorous specimen-level split dataset to achieve 95.64% accuracy in classifying foraminifera species from 2D micro-CT slices, while also providing an interactive dashboard for real-time identification and 3D matching.

The Gist

Imagine you are a detective trying to identify a suspect, but the suspect is wearing a mask and you can only see them through a series of random, blurry snapshots taken from different angles. This is the daily challenge for paleontologists (fossil hunters) studying foraminifera (tiny, single-celled marine creatures with shells).

For over a century, scientists have had to slice these tiny fossils into thin pieces and look at them under a microscope. It's like trying to figure out what a whole apple looks like by only seeing a few random slices of it. Sometimes the slice shows the core, sometimes the skin, and sometimes just a random chunk. It's hard to tell the species apart, and it takes a human expert hours to do what a machine could do in seconds.

This paper introduces ForamDeepSlice, a new "super-detective" AI that solves this problem. Here is how it works, explained simply:

1. The Problem: The "Jigsaw Puzzle" Dilemma

Foraminifera are incredibly diverse, with thousands of species. To identify them, scientists usually need to see their 3D internal structure (like the rooms inside a castle). But getting a full 3D view is expensive and slow (like hiring a team to build a full-scale model of a house just to look at one room).

Instead, scientists usually get 2D slices (like looking at one floor plan). The problem is that a single species can look completely different depending on how the slice was cut. One slice might look like a circle, another like a star, and another like a blob. Traditional AI gets confused by this, often guessing wrong.

2. The Solution: A "3D Library" of 2D Photos

The researchers didn't just take random photos. They used a high-tech 3D scanner (Micro-CT) to scan 97 actual fossil specimens and create a massive digital library.

• The Analogy: Imagine taking a real apple, scanning it in 3D, and then slicing it digitally into 109,617 different 2D pictures from every possible angle.
• They used these pictures to teach an AI. They made sure the AI learned that "this specific apple" (specimen) produced all these different slices, so it wouldn't get confused when it saw a new slice from the same apple later.

3. The Brain: Two Detectives Working Together

The researchers tried seven different types of AI "brains" (neural networks). Most were good, but two specific species—Baculogypsina and Orbitoides—were like the "master of disguise" suspects. They looked so similar to each other that even the best AI kept mixing them up.

• The Old Way (The Crowd Vote): Usually, if you have multiple AIs, you ask them all to vote. If 3 out of 5 say "It's Orbitoides," you go with that. But here, all the AIs made the same mistake on these tricky species.
• The New Way (The "PatchEnsemble"): The researchers created a smart system called ForamDeepSlice (FDS).
  • The Main Detective (ConvNeXt-Large): This is the expert who handles 95% of the cases perfectly.
  • The Specialist (EfficientNetV2-Small): This is a different expert who is specifically good at spotting the "master of disguise" species.
  • The Switch: The system works like a traffic light. If the Main Detective is unsure or thinks it's one of the tricky species, it automatically hands the case over to the Specialist. If the Main Detective is confident, it keeps the case.
  • The Result: This "conditional switching" boosted the accuracy to 95.6%, and if you look at the top 3 guesses, it's right 99.6% of the time.

4. The Dashboard: No Coding Required

Usually, using AI requires knowing how to code (Python, etc.). The researchers built a user-friendly dashboard (like a video game interface) so any geologist can use it without being a computer expert.

• Upload: You drag and drop a picture of a fossil slice.
• Identify: The AI tells you what it is and how sure it is.
• Match: It can also search a 3D library to find the exact slice orientation or similar fossils, helping scientists visualize the whole 3D shape from a single 2D picture.

Why This Matters

• Speed: What used to take a human hours can now be done in seconds.
• Accuracy: It reduces human error and fatigue.
• Real World Impact: Identifying these tiny fossils helps oil companies find new drilling sites and helps climate scientists understand how the Earth's oceans changed millions of years ago.

The Catch (Limitations)

The AI is only as good as the library it was trained on.

• It knows the 12 species in the study very well.
• If you show it a fossil species it has never seen before, it might guess wrong (or guess one of the known species incorrectly).
• It was trained on 3D-scanned slices. While it might work on regular microscope photos, that needs more testing.

In a nutshell: The researchers built a digital "fossil library," trained a smart AI to recognize tiny sea creatures from 2D slices, and created a "smart switch" system that fixes its own mistakes on the hardest cases. They then wrapped it all in a simple app so real scientists can use it to unlock the secrets of the Earth's past.

Technical Summary

Here is a detailed technical summary of the paper "ForamDeepSlice: A High-Accuracy Deep Learning Framework for Foraminifera Species Classification from 2D Micro-CT Slices."

1. Problem Statement

The classification of foraminifera (single-celled marine organisms with calcareous shells) is critical for biostratigraphy, paleoenvironmental reconstruction, and hydrocarbon exploration. Traditionally, this relies on manual identification of 2D thin sections under optical microscopes, a process that is labor-intensive and often fails to capture diagnostic internal 3D structures.
While micro-Computed Tomography (micro-CT) provides high-resolution 3D internal morphology, it creates a new bottleneck:

• Data Volume: Micro-CT generates massive volumetric datasets.
• Manual Segmentation: Separating fossils from rock matrices is slow and error-prone.
• Classification Gap: Existing deep learning methods often struggle with species-level classification due to limited datasets, lack of 3D pre-trained models, and the inherent variability of 2D slices extracted from 3D volumes (e.g., random orientations, morphological ambiguity).

The authors aim to bridge the gap between high-fidelity 3D micro-CT data and routine 2D thin-section analysis by creating an automated, high-accuracy deep learning pipeline for species classification.

2. Methodology

A. Dataset Curation and Splitting

• Source Data: 97 micro-CT scanned specimens representing 27 foraminifera species.
• Selection: 12 species were selected for the final dataset, ensuring at least four 3D models (specimens) per species to ensure robust training.
• Data Leakage Prevention: A strict specimen-level splitting protocol was employed. All 2D slices derived from a single 3D model were assigned exclusively to either the training, validation, or test set. This prevents the common error of "slice-level splitting," where adjacent slices from the same specimen appear in both training and testing, artificially inflating accuracy.
• Final Dataset: 109,617 high-quality 224×224 RGB slices (44,103 training, 14,046 validation, 51,468 testing).

B. Preprocessing and Augmentation

• Slice Extraction: Otsu's thresholding removed low-content slices; advanced segmentation isolated fossil silhouettes.
• Augmentation: To simulate the random orientations found in traditional thin sections, geometric augmentations included rotations (±45°), scaling, and flipping. Regularization techniques included CutMix and Mixup.

C. Model Architectures and Transfer Learning

• Base Models: Seven state-of-the-art 2D CNN architectures were evaluated using ImageNet pre-trained weights:
  • ConvNeXt (Base, Large)
  • EfficientNetV2 (Small, Large)
  • NASNet, MobileNet, ResNet101V2.
• Training Strategy: A two-phase approach:
  1. Backbone Freezing: 10 epochs training only the classifier head.
  2. Fine-Tuning: 20 epochs with the full network unfrozen, using mixed-precision (fp16) and early stopping.

D. The Core Innovation: PatchEnsemble (ForamDeepSlice)

Instead of standard ensemble methods (e.g., majority voting) which can propagate correlated errors, the authors developed PatchEnsemble, a confidence-gated model-switching mechanism:

• Primary Model: ConvNeXt-Large (high overall accuracy).
• Patch Model: EfficientNetV2-Small (selected for its superior recall on specific difficult classes).
• Logic:
  1. The primary model predicts a class.
  2. If the predicted class belongs to a pre-identified set of "weak classes" (specifically Baculogypsina and Orbitoides) AND the patch model exhibits higher confidence than the primary model, the system switches to the patch model's prediction.
  3. Otherwise, the primary model's prediction is retained.
• Rationale: This isolates corrections to specific failure modes without degrading performance on well-classified species.

E. Deployment

A containerized interactive dashboard (Docker-based) was developed featuring:

• Real-Time Classification: Ranked predictions with confidence scores.
• 3D Slice Matching: Similarity analysis using SSIM, NCC, Dice coefficient, and ORB feature matching to locate the input slice within 3D reference volumes.

3. Key Results

• Overall Performance: The ForamDeepSlice (FDS) ensemble achieved a test accuracy of 95.64%, a Top-3 accuracy of 99.6%, and an AUC of 0.998.
• Individual Model Comparison:
  • ConvNeXt-Large was the best single model (95.1% accuracy).
  • Standard Top-k ensembles (e.g., Top-2, Top-3 voting) failed to improve accuracy and often exacerbated errors on difficult species.
• Species-Level Analysis:
  • Most species achieved >90% F1-scores.
  • Critical Failure Modes: Baculogypsina suffered from low recall (high false negatives) due to morphological asymmetry, while Orbitoides suffered from low precision (high false positives) due to confusion with Baculogypsina and specimen damage.
  • PatchEnsemble Impact:
    • Baculogypsina recall improved from 0.243 (Top-2 ensemble) to 0.603 (FDS).
    • Orbitoides precision improved from 0.684 to 0.827.
    • Misclassifications of Baculogypsina as Orbitoides dropped from 1,302 to 952 cases.
• Cross-Modality: Preliminary tests showed the model could classify optical microscopy images (e.g., Ataxophragmium), though this is noted as anecdotal and requires further validation.

4. Key Contributions

1. Robust Dataset & Protocol: Creation of a scientifically rigorous dataset with specimen-level splitting to eliminate data leakage, setting a new standard for 3D-to-2D fossil classification benchmarks.
2. Novel Ensemble Strategy: Introduction of PatchEnsemble, a conditional switching mechanism that outperforms traditional voting ensembles by targeting specific model weaknesses rather than averaging predictions.
3. Practical Deployment: Development of a user-friendly, containerized dashboard that allows non-technical paleontologists to perform species classification and 3D slice matching without coding knowledge.
4. Bridging 3D and 2D: Demonstrating that 2D slices from 3D micro-CT scans can effectively train models for 2D thin-section identification, leveraging the rich internal morphology of CT data to solve 2D classification problems.

5. Significance and Limitations

Significance:

• Automation: Drastically reduces the time required for foraminifera identification, a bottleneck in biostratigraphy and reservoir characterization.
• Reproducibility: Provides a fully reproducible framework (code, models, and data) for the micropaleontology community.
• Accuracy: Sets a new benchmark (95.64% accuracy) for AI-assisted fossil identification, surpassing previous genus-level or lower-resolution studies.

Limitations & Future Work:

• Domain Shift: The model was trained on data from a single micro-CT facility. Performance on data from different scanners, resolutions, or reconstruction algorithms is untested.
• Homeomorphy: The model cannot distinguish between unrelated species that evolved similar shapes (homeomorphs) if they are not in the training set.
• Optical Microscopy: While preliminary cross-modality results are promising, the model is not yet validated for routine optical microscopy use.
• Data Scope: The current dataset is limited to 12 species; future work requires expanding the species count and specimen diversity.

In conclusion, ForamDeepSlice represents a significant leap forward in applying deep learning to geosciences, successfully combining advanced architectural strategies with practical tooling to automate a critical scientific workflow.