Cross-Species Adaptation of RETFound for Rodent OCT Age Estimation Reveals Strong CNN Baselines in Data-Scarce Space Biology

This paper benchmarks the cross-species transferability of the human retinal foundation model RETFound for predicting rodent age from OCT scans, finding that while the model provides scientifically useful results, a strong CNN baseline (Xception) actually outperformed it in this data-scarce space biology setting.

The Gist

The Big Idea: Can a "Human Expert" Teach a Robot to Understand Rat Eyes?

Imagine you have a world-class master chef who has spent 20 years learning exactly how to cook French cuisine. They know every subtle detail of a soufflé or a delicate sauce.

Now, imagine you take that master chef and drop them into a tiny kitchen in a remote village where they are asked to cook traditional street food from a different country. The chef isn't starting from zero—they already understand heat, salt, texture, and timing—but they aren't an expert in this new, specific menu yet.

This paper is about a "Master Chef" AI and whether it can learn to read the "menu" of a rat’s eye.

---

The Problem: The "Data Desert" of Space Biology

Scientists studying how spaceflight affects living creatures (like rats) face a huge problem: Data Scarcity.

When we send animals into space or simulate space conditions on Earth, we can’t take millions of pictures. We only have a handful. In the world of Artificial Intelligence, having only a few pictures is like trying to learn a new language by only hearing three sentences. Usually, AI needs millions of examples to get smart. Without enough data, the AI "hallucinates" or fails completely.

The Experiment: The "Human-to-Rat" Translation

The researchers wanted to see if they could cheat the "Data Desert" by using a shortcut called Transfer Learning.

1. The Expert (RETFound): They started with an AI called RETFound. This AI is like a super-genius that has already looked at 1.6 million human eye scans. It understands the "language" of eyes—layers, textures, and shapes—better than almost anything else.
2. The Task: They wanted to see if this "Human Eye Expert" could look at scans of rat eyes (using a technology called OCT, which is like a high-tech ultrasound for the eye) and accurately guess how old the rat is.
3. The Twist: They didn't just give the expert the new job; they used a technique called LoRA (think of this as giving the chef a "crash course" manual rather than making them go back to culinary school for four years).

The Results: The "Old Pro" vs. The "Specialist"

The researchers compared two different types of AI "brains":

• The Transformer (RETFound): The "Human Expert" who was trying to adapt to rats.
• The CNN (Xception): A different type of AI that is more like a "Local Specialist." It wasn't trained on human eyes; it was trained on general images (like cats, dogs, and cars) and was very good at spotting patterns.

The Surprise: Even though the "Human Expert" (RETFound) was incredibly smart, the "Local Specialist" (Xception) actually won the contest! The specialist was better at guessing the rats' ages more accurately.

Why does this matter? (The "So What?")

You might think, "If the human expert lost, is the experiment a failure?" No! It’s actually a huge win for science.

Here is why:

1. It Proves the Shortcut Works: It proved that you can take an AI trained on humans and make it useful for rats. It’s not a waste of time.
2. It Sets a "Speed Limit": It warns other scientists: "Hey, if you have a very small amount of data, don't just go for the fanciest, most complex AI. Sometimes a simpler, sturdy 'Specialist' AI will actually do a better job."
3. The "Map" is Correct: When they checked where the AI was looking, it was looking at the right parts of the eye—not just guessing randomly. It was actually "reading" the biology.

Summary in a Nutshell

Scientists tried to see if an AI trained on millions of human eyes could help study rats in space biology. While the "Human Expert" AI was very good, a simpler "Specialist" AI actually performed better. This gives future space scientists a roadmap: they now know which tools to use when they are working with the tiny, precious amounts of data that come from space research.

Technical Summary

Technical Summary: Cross-Species Adaptation of RETFound for Rodent OCT Age Estimation

1. Problem Statement

The field of space biology faces a significant bottleneck: data scarcity. High-fidelity imaging studies involving rodent spaceflight or space-analog missions are expensive and produce limited datasets. This scarcity prevents the training of deep learning models from scratch, which typically require massive amounts of labeled data to achieve robustness.

While "foundation models" (models pretrained on massive datasets) are a solution for human medical imaging, it is unknown whether a model trained on human retinal images can generalize to rodent Optical Coherence Tomography (OCT) scans. The researchers specifically investigated whether a human-centric Vision Transformer (ViT) could serve as a reliable feature extractor for predicting chronological age in rats, a critical biomarker for studying aging and physiological decay in space.

2. Methodology

The study benchmarks the transferability of RETFound—a state-of-the-art retinal foundation model pretrained on 1.6 million human retinal images—to a rodent context.

• Dataset: The researchers used the NASA Open Science Data Repository (OSD-679), consisting of OCT B-scans from Brown Norway rats.
• Task: Regression of chronological age from OCT B-scans.
• Adaptation Strategy: To prevent "catastrophic forgetting" and manage the small dataset, they used Low-Rank Adaptation (LoRA). LoRA is a Parameter-Efficient Fine-Tuning (PEFT) technique that injects trainable low-rank matrices into the transformer layers, allowing the model to adapt to new species without updating all millions of parameters.
• Experimental Design & Baselines:
  • Primary Model: RETFound + LoRA.
  • Strong CNN Baseline: An Xception model pretrained on ImageNet (to test if traditional convolutional architectures are more efficient in small-data regimes).
  • Negative Control: A Vision Transformer (ViT) trained from scratch/random initialization.
  • Validation: 3-fold rat-level cross-validation (ensuring eyes from the same rat are not split between training and testing sets to prevent data leakage).
• Evaluation Metrics: Mean Absolute Error (MAE) in days, Coefficient of Determination ($R^2$), and Inter-eye Mean Absolute Difference (MAD) for consistency.

3. Key Results

The study yielded a surprising "counter-intuitive" result regarding the superiority of foundation models:

• Performance Gap: While RETFound+LoRA was successful, the Xception CNN baseline outperformed it.
  • RETFound+LoRA: MAE $\approx$ 26.20 days; $R^2 \approx$ 0.744.
  • Xception: MAE $\approx$ 19.01 days; $R^2 \approx$ 0.853.
• Model Interpretability: Saliency maps (visualizing where the model "looks") showed that both models focused on anatomically plausible inner retinal regions, suggesting the models were learning biological features rather than imaging artifacts.
• Consistency: The models maintained high inter-eye consistency, meaning the age predictions for both eyes of a single animal were stable.

4. Key Contributions

1. Cross-Species Benchmark: It provides the first rigorous evaluation of how human-trained retinal foundation models (ViTs) transfer to rodent OCT data.
2. LoRA Implementation in Space Biology: Demonstrates the utility of Parameter-Efficient Fine-Tuning (PEFT) for adapting large models to niche, small-scale biological datasets.
3. Baseline Establishment: It establishes a "strong CNN baseline" (Xception) that future researchers in space biology should aim to beat, preventing the common pitfall of assuming a Transformer is automatically superior.

5. Significance and Implications

The paper concludes that while human retinal foundation models are "scientifically useful" and can transfer knowledge across species, they are not a "silver bullet." In data-scarce environments (like space biology), highly optimized, smaller CNN architectures may still hold an advantage over massive Transformer-based models.

This work provides a reproducible framework and a cautionary benchmark for the space biology community, guiding future efforts toward developing robust, automated biomarkers for monitoring the health of astronauts and animal models in microgravity.