Crop-OCT: a Fully Integrated Imageomics Pipeline to Identify Regional and Focal Retinopathy in Murine Models

The paper introduces Crop-OCT, an automated, end-to-end imageomics pipeline that successfully extracts and analyzes millions of features from over 20,000 OCT images across 13 murine retinopathy models to enable precise monitoring of disease progression, aging, and regional ocular heterogeneity.

The Gist

Imagine you are trying to understand how a city is changing over time. You could look at a single street corner, or you could try to look at the whole city at once. For a long time, scientists studying eye diseases in mice have been like people looking at just one street corner. They would take a picture of the eye, measure a few specific layers, and move on. But eyes are complex, and diseases often start in just one small, hidden spot before spreading.

This paper introduces a new, super-smart tool called Crop-OCT. Think of it as a high-tech, automated city inspector that doesn't just look at one street; it scans the entire city, takes thousands of photos, and uses a robot brain to find tiny cracks in the pavement that human eyes might miss.

Here is the breakdown of how it works and why it matters, using simple analogies:

1. The Problem: The "Blurry Snapshot"

Scientists use a machine called an OCT (Optical Coherence Tomography) to take pictures of a mouse's eye. It's like a super-powered ultrasound that creates a cross-section of the eye, showing the different layers (like the layers of a cake).

• The Old Way: Scientists would manually measure the thickness of the "cake layers." If the cake was uneven, they might miss a small bump or a hole because they were only looking at the center. Also, mice breathe and move, making the pictures blurry.
• The Result: They were missing the "focal" problems—small, specific areas where the disease started.

2. The Solution: The "Crop-OCT" Pipeline

The authors built a fully automated pipeline (a step-by-step assembly line) called Crop-OCT. Here is how it works, step-by-step:

• Step 1: The "Slicing" Machine (Cropping):
  Imagine the eye is a round orange. The OCT machine takes a long slice through the middle. Crop-OCT automatically cuts that slice into 8 smaller, manageable pieces (like slicing a pizza into wedges). Crucially, it keeps a map of where each slice came from (top, bottom, left, right). This preserves the "address" of every piece of the eye.

• Step 2: The "Robot Chef" (AI Segmentation):
  Once the slices are cut, a robot chef (Artificial Intelligence) looks at each piece. It doesn't just guess; it has been trained on thousands of examples to perfectly identify the different layers of the retina (the "cake layers"). It draws a line around every single layer, separating the "frosting" from the "sponge" and the "filling."

• Step 3: The "Quality Control" Gatekeeper:
  Sometimes a mouse blinks or breathes, making a picture blurry. The robot gatekeeper checks every single slice. If a slice is too blurry, it throws it away. If it's good, it keeps it. In this study, it saved about 95% of the images, which is a huge success.

• Step 4: The "Detective" (Feature Extraction):
  This is the magic part. The robot doesn't just measure thickness. It measures 267 different things for every single slice!

  • Analogy: Imagine a detective looking at a wall. A normal person just sees "a wall." The detective measures the paint thickness, the angle of the cracks, the number of bubbles, and how the texture changes.
  • Crop-OCT measures the angle of the layers (are they tilting?), the bumps (are there holes?), and the texture. It turns a picture into a massive spreadsheet of data.

3. What Did They Find?

By using this tool on over 20,000 images from 13 different types of mice with different eye diseases, they discovered things they couldn't see before:

• The "Regional" Mystery: Some diseases don't attack the whole eye at once. In one specific mouse model (the Tsc1 model), the disease was eating away the "cake" only on the top of the eye, while the bottom was fine. Without the "map" that Crop-OCT kept, they would have averaged the data and thought the eye was only slightly sick, missing the severe damage on top.
• The "Focal" Surprise: They found tiny, isolated spots of damage (focal lesions) that looked like potholes in a road. These were so small that manual checking would have missed them, but the robot found them instantly.
• The "General" Test: They tested this tool on a completely different set of mice (from a different lab) that they had never seen before. The tool worked perfectly, proving it's not just a trick for one specific mouse, but a universal tool for any eye disease.

4. Why Does This Matter?

• Speed: What used to take a human weeks to measure, the robot does in minutes.
• Precision: It finds the "needle in the haystack" (small, early-stage disease) before it becomes a "haystack" (total blindness).
• Future of Medicine: Since human eyes and mouse eyes are similar, this tool helps us understand human diseases like Diabetic Retinopathy, Macular Degeneration, and Glaucoma. It's like having a practice simulator for human eye diseases.

The Bottom Line

Crop-OCT is like upgrading from a hand-drawn map to a satellite navigation system. It doesn't just tell you that the eye is sick; it tells you exactly where, how bad it is in that specific spot, and how it's changing over time. This allows scientists to catch diseases earlier and test new drugs more accurately, potentially leading to better treatments for humans down the road.

Technical Summary

1. Problem Statement

Optical Coherence Tomography (OCT) is a standard non-invasive imaging modality used to diagnose and monitor retinopathy in both clinical and preclinical settings. While "imageomics" (using machine learning to extract biological traits from images) has advanced in fields like digital pathology and fundus imaging, its application to preclinical OCT data remains limited.

• Current Limitations: Existing methods often lack generalizability across diverse murine models, fail to preserve spatial location data (crucial for identifying regional heterogeneity), and rely on manual segmentation which is not scalable for large datasets.
• The Gap: There is no standardized, automated, end-to-end pipeline capable of segmenting retinal layers, performing quality control (QC), extracting thousands of features, and preserving the spatial context of thousands of OCT images across diverse genetic models of retinal disease.

2. Methodology: The Crop-OCT Pipeline

The authors developed Crop-OCT, a fully integrated, automated pipeline designed to process preclinical OCT data from mouse models. The workflow consists of the following stages:

• Data Acquisition & Preprocessing:

  • Cohort: 13 genetic models of retinopathy (including Retinitis Pigmentosa, Leber Congenital Amaurosis, Diabetic Retinopathy, AMD, etc.) and 2 inbred strains, totaling 336 mice imaged from 3 to 27 months of age.
  • Imaging: Paired fundus and OCT images acquired using a Phoenix MICRON IV system. Images were taken at three locations (superior, midline, inferior) to capture spatial heterogeneity.
  • Cropping: To account for ocular curvature, each OCT image was cropped into three regions perpendicular to the midline. This preserved the relative $(x, y)$ spatial coordinates of each crop relative to the optic nerve head.

• Segmentation & AI Models:

  • Tissue Segmentation: An Ilastik model was trained to segment the tissue (retina, RPE, choroid) from the background and define the midline.
  • Layer Segmentation: Two Swin-UMamba models (based on the Mamba architecture) were trained using manual ground truth annotations.
    • Model 1: Segmented 8 distinct retinal layers in cropped OCT images.
    • Model 2: Segmented the optic nerve head in fundus images.
  • Training Data: 296 cropped OCT images and 219 fundus images were manually annotated for training.

• Feature Extraction & Quantification:

  • Metrics: The pipeline extracted 267 features per image, including layer thickness, local thickness statistics (mean, SD, range, IQR), and normalized thickness.
  • Spatial Analysis: By linking crop coordinates to the optic nerve, the pipeline mapped features to specific anatomic locations, enabling the detection of regional heterogeneity.
  • Focal Lesion Detection: Novel algorithms were developed to detect focal disruptions:
    • Skeletonization: Generated medial axes for each layer.
    • Angle Analysis: Calculated the tangent angle of the medial axis in a sliding window to detect inflections, domain breaks, and bifurcations.
    • Detachment Mask: Added a specific segmentation mask for retinal detachments.

• Quality Control (QC) & Statistics:

  • QC Thresholds: Images were filtered based on domain counts, bifurcation numbers, and layer continuity. 94.5% of images passed QC.
  • Statistical Analysis: Used non-parametric tests (Wilcoxon rank-sum, Kruskal-Wallis) with FDR correction. Unsupervised hierarchical clustering was used to group images by phenotype (e.g., focal lesions vs. normal).

3. Key Contributions

• End-to-End Automation: Created a scalable pipeline processing >20,000 OCT images and extracting nearly 6 million data points.
• Spatial Preservation: Unlike previous methods that averaged data, Crop-OCT preserves the $(x, y)$ location of every crop, allowing for the mapping of regional phenotypes (e.g., superior vs. inferior retinal thinning).
• Generalizability: The pipeline was trained on a diverse set of models but successfully generalized to an independent, blinded external dataset (NZO and WSB/EiJ mice) without retraining, demonstrating robustness across different genetic backgrounds and disease mechanisms.
• Focal Lesion Detection: Developed a novel method using skeleton angle analysis and bifurcation detection to identify focal disruptions (e.g., retinal detachments, localized degeneration) that are often missed by global layer thickness measurements.

4. Key Results

• Disease Progression Monitoring: The pipeline successfully quantified progressive thinning of retinal layers (ONL, IS/OS, IPL) in models like Rpe65rd12/rd12 and Ins2Akita/+, correlating strongly with histopathology. It also identified sex-specific phenotypes in diabetic mice (males showed significant thinning; females did not).
• Regional Heterogeneity: In the Tsc1 mouse model (a model for AMD), the pipeline detected significant regional heterogeneity in Outer Nuclear Layer (ONL) thinning, specifically in the superior region of the eye, which was validated by histology. This highlighted the necessity of spatial mapping.
• Focal Lesion Identification:
  • Retinal Detachments: Successfully clustered Cep290rd16/rd16 mice based on retinal detachment features.
  • Focal Degeneration: Identified a subset of WSB/EiJ mice in the external dataset with focal RPE infiltration and localized degeneration that was not present in the training set, proving the tool's ability to discover novel phenotypes.
• Validation: Results were validated against manual histopathology (H&E and immunohistochemistry), showing high concordance in relative layer thickness trends, with minor absolute differences attributed to tissue fixation artifacts.

5. Significance

• Unified Framework: Crop-OCT provides a unified standard for feature extraction in preclinical retinal disease, facilitating cross-study comparisons and meta-analyses.
• AI Foundation: The pipeline serves as a foundational tool for future multimodal data integration and AI applications in "oculomics" (using eye imaging to diagnose systemic diseases like Alzheimer's or diabetes).
• Scalability: By automating the analysis of thousands of images, it accelerates drug discovery and the characterization of genetic models, moving beyond small-scale manual studies to high-throughput imageomics.
• Open Source: The authors have released the code, Singularity container, and data on GitHub and Zenodo, promoting reproducibility and adoption in the broader scientific community.

In summary, Crop-OCT represents a significant leap forward in preclinical ophthalmic research, transforming raw OCT data into high-dimensional, spatially resolved biological insights capable of detecting subtle, regional, and focal disease phenotypes.