📄 cancer biology

Correlation of OCT-Based Radiomic Signatures With Dose-Associated Radiation Response in Tumor Spheroids

This study demonstrates that optical coherence tomography (OCT) radiomics provides a sensitive, reproducible, and label-free high-throughput method for quantifying radiation dose responses in tumor spheroids, significantly outperforming conventional brightfield morphology without requiring concurrent brightfield imaging.

Original authors: Arndt, M. D., Hansler, R., Tirinato, L., Tkachenko, A., Seco, J., Schepers, U., Spadea, M. F.

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arndt, M. D., Hansler, R., Tirinato, L., Tkachenko, A., Seco, J., Schepers, U., Spadea, M. F.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: A New Way to "X-Ray" Tiny Tumor Balls

Imagine you are a scientist trying to figure out how well a new radiation treatment works. In the past, you might have grown tiny, 3D balls of cancer cells (called tumor spheroids) in a lab dish, zapped them with radiation, and then looked at them under a regular microscope.

The problem with the regular microscope is that it's like looking at a globe through a flat window. You can see the outline of the globe, but you can't see what's happening inside the continents or oceans. If the radiation kills cells in the middle of the ball, a flat image might not show it clearly.

This paper introduces a new tool called OCT (Optical Coherence Tomography). Think of OCT as a 3D ultrasound for light. Instead of just seeing the flat outline, it takes a "slice-by-slice" picture of the entire ball, allowing scientists to see deep inside without cutting it open or using any dyes.

The Experiment: The "Radiation Test"

The researchers grew two sizes of these tumor balls (small ones with 5,000 cells and big ones with 10,000 cells). They then gave them different doses of radiation, ranging from "no radiation" (0 Gy) to a heavy dose (12 Gy).

They wanted to answer one main question: Can we use a computer to look at these 3D light pictures and accurately guess how much radiation the tumor ball received, just by looking at how the ball changed?

The Two Competitors: The "Flat Photo" vs. The "3D Scan"

To test this, they set up a race between two ways of looking at the data:

The Old Way (Brightfield): This is like taking a standard 2D photograph of the tumor ball. The computer measures the ball's size and shape.
- The Flaw: It's like trying to guess how much a water balloon has been squeezed just by looking at its shadow. If the balloon gets a little squishy inside but the outside shape doesn't change much, the photo misses the damage.
The New Way (OCT Radiomics): This is the 3D light scan. But instead of just looking at the picture, they used a special computer program (called Radiomics) to turn the picture into a long list of numbers. These numbers describe things like:
- How rough the surface is.
- How "bumpy" the inside texture is.
- How the cells inside are moving or vibrating (using a technique called Speckle Variance).

The Results: The 3D Scan Wins

The researchers fed these lists of numbers into different types of AI models (like a team of different detectives) to see who could best guess the radiation dose.

The Winner: The OCT-only team (the 3D scan detectives) was much better at guessing the dose than the Brightfield-only team (the 2D photo detectives).
- The Score: The 3D scan team got a score of roughly 0.8 out of 1.0, while the 2D photo team only got about 0.6.
The Surprise: The researchers thought that combining the 2D photo with the 3D scan would make the AI even smarter (like having two detectives working together). However, adding the 2D photo didn't really help. The 3D scan alone was already so good that the extra photo was mostly redundant.

What Did the Computer Actually "See"?

Why was the 3D scan so much better? The paper found that the radiation caused subtle changes inside the tumor ball that the flat photo couldn't see:

Roughness: As the radiation dose went up, the surface of the tumor ball became more jagged and irregular.
Internal Chaos: The inside of the ball became "patchy" and messy. The computer detected that the cells inside were no longer moving or vibrating in a smooth, organized way; instead, the movement became scattered and chaotic.

The computer learned to spot these "internal messiness" patterns and realized: "Ah, this specific type of internal chaos means the tumor got a high dose of radiation."

The Bottom Line

This study proves that you don't need to look at a tumor ball from the outside to know how it's reacting to radiation. By using a special 3D light scanner and a smart computer program, scientists can get a highly accurate, repeatable, and label-free (no dyes needed) measurement of how well the radiation is working.

It's like realizing that to understand how a cake is baking, you don't just need to look at the crust (the 2D photo); you need to be able to see the rising and bubbling inside the batter (the 3D scan). The paper shows that looking inside is the key to understanding the treatment.

Technical Summary: Correlation of OCT-Based Radiomic Signatures With Dose-Associated Radiation Response in Tumor Spheroids

Problem Statement
Three-dimensional tumor spheroids serve as a critical bridge between 2D cell cultures and in vivo tissue for radiobiology studies, accurately reproducing cell–cell interactions and diffusion-driven gradients (e.g., oxygen, nutrients) that influence radiotherapy response. However, a significant gap exists in scalable, reproducible readouts for quantifying dose-dependent radiation response. Conventional brightfield (BF) imaging provides only 2D projections, failing to resolve internal 3D structures or depth-dependent changes. Fluorescence microscopy, while specific, often requires exogenous probes that can perturb live samples and is frequently limited to endpoint measurements. Confocal microscopy offers optical sectioning but is time-intensive and suffers from signal loss in thick samples. Consequently, there is a need for a label-free, non-destructive, high-throughput method capable of capturing the complex, three-dimensional, and dynamic alterations induced by irradiation.

Methodology
This in vitro, cross-sectional study utilized SAS human tongue squamous cell carcinoma spheroids seeded at two densities (5,000 and 10,000 cells) and irradiated with doses ranging from 0 to 12 Gy. Imaging was conducted on days 1, 4, 6, 8, and 11 post-irradiation.

Imaging Modalities:
- Brightfield: Used to measure 2D morphological metrics (diameter and projected area) manually via ImageJ to serve as a baseline.
- Optical Coherence Tomography (OCT): A spectral-domain OCT system (880 nm center wavelength) acquired depth-resolved volumetric data. Each acquisition yielded two co-registered volumes: a structural intensity volume and a speckle-variance (SV) volume. The SV volume was computed from 20 repeated B-scan frames to capture temporal intensity fluctuations related to cellular dynamics and microenvironmental activity.
Radiomics Pipeline:
- Segmentation: OCT volumes were cropped and masked using Otsu thresholding and morphological operations to isolate spheroid tissue.
- Feature Extraction: Using Radiomics.jl, features were extracted from both intensity and SV volumes, including shape (from intensity only), first-order statistics, and texture descriptors (GLCM, GLRLM, GLSZM, GLDM).
- Robustness Filtering: A repeatability experiment (six scans of a single spheroid) was used to filter features with an interquartile range-to-median ratio (IQR/median) exceeding 35%, removing 24 of 198 initial features.
- Feature Selection: Features underwent correlation pruning (Spearman $|\rho| \ge 0.95$ ) and ensemble ranking based on permutation importance and univariate scores. Selection was performed independently for "OCT-only" and "Combined OCT+BF" feature sets to ensure fair comparison.
Modeling: Five regression models (Random Forest, Extra Trees, SVR, LightGBM, XGBoost) were trained to predict irradiation dose. Performance was evaluated using repeated 5-fold cross-validation, comparing BF-only, OCT-only, and combined feature sets.

Key Contributions

Validation of OCT Radiomics: The study demonstrates that OCT-derived radiomic features (structural intensity and speckle variance) can quantitatively predict radiation dose in tumor spheroids with higher accuracy than conventional brightfield morphology.
Label-Free High-Throughput Readout: It establishes a pipeline that does not require exogenous probes or destructive sampling, enabling longitudinal assessment of spheroid response.
Speckle Variance Utility: The work highlights the value of speckle-variance contrast, showing that temporal fluctuations in OCT signals provide complementary information to structural imaging regarding treatment-induced heterogeneity.
Reproducibility Analysis: The study provides empirical evidence that selected radiomic features exhibit low variability across repeated acquisitions, confirming that observed cohort-level variability reflects biological differences rather than measurement noise.

Results

Performance Comparison: OCT-only models consistently outperformed the brightfield-only baseline across all regressors. The median $R^2$ for OCT-only models ranged from 0.77 to 0.85, compared to 0.61 to 0.69 for BF-only models ( $p < 0.001$ for all comparisons).
Impact of Multimodal Fusion: Adding brightfield features to OCT did not yield a consistent performance benefit. The combined OCT+BF models achieved median $R^2$ values of 0.80 to 0.84. The improvement over OCT-only was statistically significant only for the Random Forest model ( $p = 0.026$ ) with low power (0.62), and non-significant for all other models. Notably, for LightGBM, the OCT-only model achieved a marginally higher median $R^2$ (0.85) than the combined model (0.84).
Feature Robustness: Four key features (Brightfield Area, Intensity GLRLM GrayLevelNonUniformity, Shape Sphericity, and Variance GLDM GrayLevelNonUniformity) were consistently ranked highly. Repeatability testing showed low relative dispersion (IQR/median = 0.032–0.047) across six repeated scans, while cohort-level variability was orders of magnitude higher, confirming the features' discriminative power.
Visual Correlation: Qualitative analysis of OCT en face projections revealed that higher doses (8–12 Gy) induced increased boundary irregularity and a shift from smooth, concentric SV patterns to heterogeneous, patchy distributions, which were captured by the selected texture features.

Significance and Claims
The paper claims that OCT radiomics provides a sensitive, reproducible, and label-free high-throughput readout for assessing spheroid radiation dose response. A primary conclusion is that OCT-only analysis outperforms the current standard of brightfield-based morphology, suggesting that depth-resolved structural and dynamic information encodes substantial dose-related signals that 2D projections miss.

The authors modestly claim that while a small set of shared predictors combining brightfield size dynamics with OCT texture and shape features emerged, the addition of brightfield data generally offers no consistent practical benefit over OCT alone. This supports the feasibility of an OCT-only pipeline for dose-correlation modeling, eliminating the need for concurrent brightfield acquisition or manual annotation. The study acknowledges limitations, including the use of a single cell line, a cross-sectional design (preventing true within-spheroid longitudinal tracking), and the indirect nature of dose correlation as a surrogate for biological endpoints. The authors suggest future work should focus on enabling true longitudinal OCT acquisition, validating across diverse cell lines, and linking radiomic signatures to orthogonal biological endpoints.