📄 bioengineering

LIT (Layer-Wise Image Trajectories): In Situ Monitoring for Early Quality Prediction and Anomaly Detection in Acellular and Cell-Laden Two-Photon Polymerization

This paper introduces Layer-wise Image Trajectories (LIT), an in situ monitoring method that analyzes coaxial image comparisons during two-photon polymerization to predict fabrication quality and detect anomalies early in the build process for both acellular and cell-laden hydrogels, overcoming the limitations of traditional post-processing evaluation.

Original authors: Prioglio, E., Scrocciolani, C., Colosimo, B. M.

Published 2026-08-17

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Prioglio, E., Scrocciolani, C., Colosimo, B. M.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine trying to build a tiny, three-dimensional city out of jelly, but this jelly is alive. You want to construct it with such precision that every street and building is the size of a single cell, creating a perfect home for living tissue. This is the world of Two-Photon Polymerization (2PP). Think of it as a super-precise laser pen that can draw inside a drop of liquid, hardening it only where the laser touches, layer by layer. It's the ultimate tool for making "bio-inks"—gels that hold living cells—to create artificial tissues for medicine or research.

However, there's a catch. Unlike regular plastic resin, these living gels are incredibly sensitive. They are like a house of cards in a windstorm; tiny changes in the laser's power or speed can cause the whole thing to collapse, or worse, look perfect at first but fall apart later. The biggest problem? You usually don't know if your tiny city survived until you wash away the liquid and look at the ruins. By then, the living cells might be dead, and the experiment is a failure. Scientists have been trying to peek at the construction process while it's happening, but it's been like trying to judge a cake's doneness just by looking at a single, blurry photo of the batter.

This paper introduces a clever new way to watch the construction happen in real-time, called Layer-wise Image Trajectories (LIT). Instead of trying to guess if a single layer is "good" or "bad" (which is often impossible because the images are fuzzy), the researchers taught a computer to compare layers against each other, like a judge comparing two runners in a race to see who is ahead. By tracking how the "score" of the printing process changes from the very first layer to the hundredth, they created a unique "fingerprint" for every print.

Here is the magic: They found that these fingerprints tell the story of the print's fate long before it's finished. If the score curve stays flat, the print is failing. If it shoots up too fast, the structure is getting damaged. If it follows a smooth, steady path, the structure is likely to succeed. Remarkably, this works even when living cells are mixed into the ink. The cells act like tiny, unpredictable obstacles that scatter the laser light, making the process even more chaotic. Yet, the LIT system could spot these chaotic prints early on, distinguishing between a successful build, a collapsed one, and one that is overcooked, all while the printer was still running.

The researchers tested this on 48 different prints, some with cells and some without, using various laser settings. They discovered that simply setting the machine to the "right" numbers wasn't enough; sometimes, even with the same settings, one print would fail while its twin succeeded. This proved that the outcome depends on what actually happens inside the machine during that specific print, not just the settings on the screen. By comparing the live "score" of a new print against a "control band" built from known successful prints, the system could raise an alarm as early as the 11th layer if something was going wrong. This means that in the future, scientists might be able to stop a failing print halfway through, saving precious time, expensive materials, and most importantly, the living cells inside. It's like having a smart assistant that whispers, "Hey, this cake is burning," while it's still in the oven, rather than waiting until you pull out a charred mess.

Technical Summary: LIT (Layer-Wise Image Trajectories) for In Situ Monitoring in Two-Photon Polymerization

Problem Statement
Two-photon polymerization (2PP) is a premier technology for fabricating cell-scale, 3D tissue microenvironments. However, the transition from conventional photoresists to hydrogel-based bioinks introduces significant unpredictability. This issue is exacerbated when living cells are embedded within the resin; cells scatter and absorb laser light, creating a spatially stochastic local energy landscape that cannot be controlled at the parameter level. Consequently, identical nominal process parameters (laser power and scan speed) often yield vastly different outcomes across replicates, ranging from complete fabrication failure to structural integrity or damage. Furthermore, traditional quality assessment relies on ex situ evaluation after development (washing away unpolymerized resin), a stage that is too late for corrective action. Existing in situ monitoring approaches often rely on discrete quality labels or single-frame analysis, which fail to capture the continuous progression of polymerization or the cumulative history of the build required to predict final dimensional stability.

Methodology
The authors propose a framework for in situ monitoring that tracks the continuous evolution of the fabrication state through a Layer-wise Image Trajectory (LIT). The methodology involves the following steps:

Experimental Setup: A full-factorial design of experiments (DOE) was conducted using a commercial 2PP system (Nanoscribe Quantum X shape) on two formulations: an acellular hydrogel (15% PhotoGel95%) and a cell-laden variant (containing C2C12 murine myoblasts). Laser power (110–140 mW) and scan speed (10,000–70,000 µm/s) were varied.
Data Acquisition: Coaxial optical images were captured at the completion of each of the 100 layers per print. These images suffer from low intrinsic contrast and artifacts (out-of-focus signals, cell scattering), making direct segmentation or absolute classification of polymerization states unreliable.
Model Architecture (Siamese Network): To overcome the limitations of absolute labeling, the authors employed a Siamese neural network (ResNet18 backbone) trained on pairwise comparisons of same-layer images from different prints.
- Labeling Strategy: Based on psychophysics and human preference theory, experts assigned ranking labels ( $y \in \{-1, 0, +1\}$ ) indicating which of two same-layer images exhibited more advanced polymerization (based on contrast, boundary sharpness, and pore definition).
- Training Objective: The network minimizes a margin-based hinge loss to learn a scalar embedding representing the "polymerization score." Crucially, the model was trained without using post-processing outcome labels, relying solely on visual relative judgments.
Trajectory Construction: The trained encoder maps each of the 100 layers of a print to a scalar score. These scores are smoothed using an Exponential Moving Average (EMA) to generate the LIT curve for each print.
Anomaly Detection: A reference band (Upper and Lower Control Limits, UCL/LCL) is established using the LIT curves of prints confirmed to have structural integrity. New prints are monitored in real-time; if a smoothed score deviates beyond the control limits (calculated with Bonferroni correction for multiple comparisons), an alarm is triggered.

Key Results

Construct-Specific Variability: The study confirmed that fabrication outcomes are not solely determined by process parameters. Identical parameter settings produced mixed outcomes (Classes 0–3) across replicates, driven by the unique, stochastic interaction of cells and resin during each specific print.
Effect of Cells: The presence of cells shifted the process window, allowing for successful construct formation at lower energy doses compared to acellular formulations. However, this also increased the risk of over-polymerization and damage (burn marks) at higher doses.
LIT Performance: The LIT curves successfully clustered according to post-processing outcomes (Class 0: no structure; Class 1: collapse; Class 2: integrity; Class 3: damage) without any class supervision during training.
- Acellular: Class 2 prints showed elevated scores from the start; Class 0 prints remained flat.
- Cell-laden: Class 3 (damaged) prints showed a steep, early rise in scores, while Class 0 prints remained flat.
Early Detection: The monitoring framework detected anomalies at early-to-mid layers (e.g., layer 11 for over-polymerization, layer 44 for under-polymerization) well before the build was complete or development occurred. This allows for the identification of prints that will fail or require intervention.

Significance and Claims
The paper claims to present the first application of in situ quality prediction and anomaly detection specifically to cell-laden two-photon polymerization.

The authors emphasize two distinct advantages of their approach over existing literature:

Continuity vs. Discreteness: Unlike classification models that assign discrete labels, the LIT provides a continuous score that preserves the gradations of fabrication state.
History-Awareness: The trajectory encodes the rate and consistency of polymerization evolution, capturing information that single-layer assessments miss.

The study concludes that process parameters are insufficient predictors of quality in hydrogel-based 2PP, particularly with living cells. The proposed LIT framework offers a non-invasive, data-driven method to monitor the "what occurred" during each individual print, enabling early detection of anomalous behavior and shifting quality assessment from a post-hoc evaluation to a real-time, actionable signal. The authors note that while the framework is promising, future work requires expanding the dataset across diverse geometries and bioinks to refine control limits and characterize transition regions between partial and full polymerization.

More like this