How to gain valuable insight from scarce data with Machine Learning: a post-hoc explanation tool to identify biases in biological images classification

This study demonstrates that applying post-hoc explanation tools like SHAP to scarce biological image datasets can reveal hidden biases, such as models overfitting to individual animal characteristics rather than biological patterns, while simultaneously uncovering how to reframe tasks to extract valid biological insights from limited data.

The Gist

Imagine you are a detective trying to solve a mystery: Can you tell the difference between a wound that is healing perfectly (regenerating) and one that is healing with a scar?

In the world of biology, scientists have a powerful new tool for this: Machine Learning (ML). Think of ML as a super-smart, hyper-observant intern who can look at thousands of photos of mouse tissue and learn to spot patterns faster than any human.

However, there's a catch. In biology, taking these photos is expensive, time-consuming, and ethically tricky. You can't just take a million photos of a million mice. You only have a tiny dataset—maybe photos from just 28 mice.

Here is the story of what happened when the scientists tried to use their super-smart intern on this tiny dataset, and how they used a special "flashlight" to find the truth.

1. The Trap: The Intern Cheated

The scientists gave their AI intern the task: "Look at these photos and tell me if this mouse is healing with a scar or regenerating."

The intern studied the training photos and seemed to ace the test. It got 100% correct! The scientists were thrilled. But then, they showed the intern new photos from mice it had never seen before.

The intern failed miserably. It got it wrong every time.

Why?
It turns out the intern wasn't actually learning about "scars" vs. "regeneration." It was cheating. It had learned to recognize individual mice.

Think of it like this: Imagine you are trying to teach a child to distinguish between "Apples" and "Oranges." But, by pure coincidence, every time you showed an Apple, it was held by Bob, and every time you showed an Orange, it was held by Alice.
The child learns the pattern: "If Bob is holding it, it's an Apple. If Alice is holding it, it's an Orange."
The child isn't learning about fruit; they are learning about Bob and Alice. When you give them a picture of an Apple held by Alice, they get confused and guess wrong.

That's exactly what happened here. The AI learned the unique "fingerprint" of each specific mouse (maybe a tiny speck of dust, a specific angle, or a unique biological quirk) rather than the actual healing process.

2. The Flashlight: SHAP (The Post-Hoc Explanation)

The scientists didn't just give up. They used a tool called SHAP. Imagine SHAP as a magnifying glass or a flashlight that shines on the AI's brain to see exactly what it was looking at when it made a decision.

When they shined this light on the AI, they saw something fascinating:

• The features the AI used to guess "Regeneration" were the exact same features it used to guess "This is Mouse #5."
• The AI was solving the wrong puzzle. It was playing "Guess Who?" instead of "Diagnose the Wound."

3. The Twist: Finding the Hidden Treasure

But the story doesn't end with failure. The scientists looked closer at the AI's mistakes.

Even though the AI couldn't tell "Scar" from "Regeneration," it was very good at grouping the mice. When they looked at the confusion matrix (a chart of where the AI got confused), they noticed a pattern:

• The AI often confused a mouse from Day 3 with another mouse from Day 3.
• It confused a mouse from Day 10 with another mouse from Day 10.

It was as if the AI was saying: "I can't tell if this is a scar or a regenerating tissue, but I can definitely tell if this photo was taken 3 days after the injury or 10 days after!"

The Breakthrough:
The scientists realized the data did contain valuable information, just not the information they originally asked for. The biological changes between Day 3 and Day 10 were much stronger and easier to see than the subtle differences between scarring and regenerating.

So, they changed the question. Instead of asking, "Is this a scar or regeneration?" they asked, "Is this Day 3 or Day 10?"

Result: The AI became a genius again. It successfully learned to distinguish the days.

The Big Lesson

This paper teaches us a vital lesson about using AI in science, especially when data is scarce:

1. AI is a mirror: It will find patterns, even if those patterns are accidental (like recognizing Bob vs. Alice instead of Apples vs. Oranges).
2. Don't just trust the score: Just because an AI gets a high score on a test doesn't mean it learned what you think it learned.
3. Look behind the curtain: By using tools like SHAP to "explain" the AI, scientists can catch these biases.
4. Pivot when necessary: Sometimes, the data can't answer your original question, but it can answer a different, equally important one. In this case, the AI couldn't predict the final outcome, but it could tell us exactly how far along the healing process was.

In short: The scientists used a "flashlight" to realize their AI was cheating by recognizing individual mice. Once they caught it, they realized the AI was actually a master of tracking time, not tissue types. This turned a failure into a discovery, proving that even with tiny, imperfect datasets, we can still find valuable biological insights if we know how to look.

Technical Summary

1. Problem Statement

The study addresses a critical challenge in biomedical research: the application of Machine Learning (ML) to scarce datasets. In fields like regenerative medicine, data collection is constrained by ethical considerations, cost, and the time required to manipulate living organisms (e.g., mice).

• The Specific Task: The authors aimed to use Deep Learning (DL) and traditional ML to classify microscopic images of mouse adipose tissue to distinguish between two healing outcomes: regeneration (tissue repair) and scarring.
• The Core Issue: Standard ML models often fail to generalize in these settings due to spurious correlations. Instead of learning biologically relevant patterns of tissue repair, models may inadvertently learn to recognize specific individuals (mice) or other dataset-specific artifacts, leading to high training accuracy but failure on independent test sets.
• The Gap: Many studies in biomedical imaging fail to follow ML best practices (e.g., proper data leakage prevention, rigorous validation), leading to false conclusions about model efficacy.

2. Methodology

The researchers employed a multi-stage approach combining standard classification, individual identification, and post-hoc explainability to diagnose model behavior.

A. Data Collection and Preprocessing

• Source: Images were acquired from 28 C57BL/6 male mice using a preclinical model where tissue repair was directed toward either regeneration (via naloxone treatment) or scarring (via small resection).
• Imaging: Confocal biphoton microscopy captured 3D tiles (composed of 2D slices) of tissue at Day 3 (D3) and Day 10 (D10) post-injury. Two channels were used: nuclei (DRAQ5 stain) and collagen fibers (Second-Harmonic Generation).
• Dataset Size: 48,788 2D images total.
• Preprocessing: Images were cropped (removing dark borders), normalized (0–1), and filtered to remove noise. Crucially, all 2D slices from a single 3D tile were kept together to prevent data leakage.

B. Classification Tasks

The study executed two distinct classification tasks using the same dataset:

1. Binary Classification (Original Goal): Distinguishing Regenerating vs. Scarring tissues.
   • Split Strategy: Images from different mice were strictly separated into training and test sets to ensure the model could not rely on individual mouse identity.
2. Individual Classification (Diagnostic Goal): Distinguishing Mouse A vs. Mouse B vs. ... Mouse N.
   • Split Strategy: Images from the same mouse were split between training and test sets (90/10 split) to verify if the model could learn individual-specific signatures.

C. Models and Algorithms

• Deep Learning: Three Convolutional Neural Networks (CNNs): ResNet18, SqueezeNet, and VGG11.
• Traditional ML: XGBoost trained on features extracted via the PyFeats library (2682 features reduced to ~900-1000 via Boruta selection).
• Explainability Tool: SHAP (SHapley Additive exPlanations) was used to analyze feature importance and dependence plots.

D. Analysis Strategy

• Error Analysis: Examining confusion matrices to identify non-random error patterns.
• Feature Comparison: Comparing the top features used by the "Binary" model vs. the "Individual" model using SHAP rankings and dependence plots.
• Subgroup Testing: Training models on specific combinations of subgroups (e.g., D3 vs. D10) to isolate which biological variables the models were actually learning.

3. Key Results

A. Failure of Binary Classification (Regeneration vs. Scarring)

• Training Performance: Models achieved high accuracy (0.75–0.97) on training and validation sets.
• Test Performance: Models failed completely on independent test sets, dropping to ~0.5 accuracy (random guessing).
• Conclusion: The models did not learn the biological difference between regeneration and scarring; they overfitted to the specific training distribution.

B. Success of Individual Classification

• Performance: Models successfully distinguished individual mice on independent test sets, achieving accuracies of 0.73–0.89 (significantly higher than the random chance of ~0.05).
• Implication: The images contained strong, learnable "signatures" unique to each mouse, which the models prioritized over the tissue repair status.

C. SHAP Analysis Reveals the Bias

• Shared Features: The SHAP analysis revealed that the top features used to classify "Regeneration vs. Scarring" were almost identical (77% overlap) to those used to classify "Mouse Identity."
• Dependence Plots: The dependence plots for the binary "Regenerating" class closely mirrored the superposition of dependence plots for the specific mice that happened to be in the "Regenerating" group in the training set.
• Diagnosis: The models were solving the task by recognizing which mouse the image came from, rather than the healing state of the tissue.

D. Discovery of Learnable Biological Signal

• Error Patterns: When the models misclassified mice, the errors were not random. They clustered by timepoint (D3 vs. D10) and healing outcome.
• Hypothesis Validation: The authors hypothesized that the time elapsed since injury (D3 vs. D10) was a stronger biological signal than the healing outcome.
• Confirmation: When the task was changed to classify D3 vs. D10, the models successfully generalized to independent test sets. This confirmed that the dataset contained valid biological information, but the original task (Regeneration vs. Scarring) was too subtle for the limited data, causing the model to latch onto the stronger "Individual" and "Time" signals.

4. Key Contributions

1. Methodological Framework for Bias Detection: The paper proposes a novel workflow using post-hoc explanation tools (SHAP) combined with individual identification tasks to diagnose why a model fails to generalize. Instead of discarding a failing model, the authors used it to uncover hidden biases.
2. Demonstration of Spurious Correlations: It provides a concrete case study showing how models in scarce-data regimes can learn "who" (the individual) rather than "what" (the biological condition), leading to deceptively high training scores but zero generalization.
3. Extraction of Value from Imperfect Data: The study demonstrates that even when a primary hypothesis (Regeneration vs. Scarring) cannot be validated due to data scarcity, secondary biological insights (e.g., distinguishing D3 from D10) can be extracted by analyzing model errors and feature dependencies.
4. Critique of Current Practices: It highlights the prevalence of data leakage and poor validation in biomedical ML literature, urging researchers to rigorously test generalization on independent individuals rather than just independent image patches.

5. Significance

This paper is significant for the intersection of Machine Learning and Biomedical Research:

• For Biologists: It offers a cautionary tale and a practical toolkit. It shows that "black box" models can be "opened" to reveal if they are learning biology or just memorizing experimental artifacts.
• For ML Practitioners: It validates the use of explainable AI (XAI) not just for trust, but as a diagnostic tool for dataset quality and task feasibility.
• For Regenerative Medicine: It suggests that while early prediction of regeneration vs. scarring is difficult with current image data, the temporal dynamics (D3 vs. D10) are robustly detectable. This shifts the focus of future research toward time-series analysis or larger datasets to capture the subtler regeneration signal.

In summary, the authors turned a "failed" classification experiment into a successful discovery of data structure, proving that careful examination of model explanations can transform scarce, biased data into actionable biological insight.