💻 computer science

Structured Proxy Features for Multimodal NSCLC Survival Prediction from Pretreatment CT

This paper demonstrates that augmenting multimodal NSCLC survival prediction with structured proxy features derived from a radiomic-parameterized cellular automaton to capture tumor heterogeneity-morphology interactions significantly improves performance on the Lung1 cohort compared to conventional radiomic and deep learning approaches.

Original authors: Huu Phong Nguyen, Delower Hossain, Ehsan Saghapour, Zhandos Sembay, Jake Y. Chen

Published 2026-08-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Huu Phong Nguyen, Delower Hossain, Ehsan Saghapour, Zhandos Sembay, Jake Y. Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future of a very complex, living city inside a person's body: a tumor. For decades, doctors and scientists have tried to guess how fast this city will grow or how long a patient might survive by looking at a few key stats, like the size of the city or the age of the people living there. They also use advanced cameras (CT scans) to take pictures of the city. But here's the catch: a tumor isn't just a static lump; it's a chaotic, shifting landscape where different parts grow at different speeds, some areas die off, and the edges are often jagged and messy. Traditional methods often treat the tumor like a simple box of data, counting up features as if they were independent items on a grocery list. They miss the big picture of how these messy parts interact with each other.

This is where a new kind of detective work comes in. Scientists are now trying to use "radiomics," which is like turning a picture into thousands of tiny numbers to describe the tumor's texture and shape, and "deep learning," where computers learn to see patterns on their own. But even these smart tools sometimes struggle to see the hidden story of how the tumor's chaos and shape work together to determine a patient's fate. The big question is: Can we teach a computer to not just see the tumor, but to simulate how it might behave, just by looking at a single snapshot? If we can build a tiny, virtual model of the tumor's growth based on its picture, could that help us predict who is at higher risk?

The Paper's Big Idea: Building a Virtual "What-If" Machine

In this paper, a team of researchers from the University of Alabama at Birmingham decided to try something different. Instead of just letting a computer stare at a CT scan and guess, they built a "virtual simulator" right into their prediction model. Think of it like this: if a standard AI is like a student memorizing a map of a city, this new approach is like giving the student a toy set and asking them to build the city themselves, then watching how it grows and changes over time.

The researchers focused on Non-Small Cell Lung Cancer (NSCLC), which is the most common type of lung cancer. They took a public database of 390 patients (called the "Lung1" cohort) who had CT scans taken before their treatment started. Their goal was to see if adding these "simulated" insights could help predict survival better than the best methods used so far.

How They Built the Simulator

The team created a clever trick called a "radiomic-parameterized cellular automaton." That sounds like a mouthful, but here's the simple version:

The Rules: They set up a virtual grid (like a 3D checkerboard) representing the tumor. They gave the grid simple rules for how cells behave: some grow, some die, and some turn into aggressive "bad" cells.
The Inputs: Instead of guessing the rules, they looked at the patient's actual CT scan. They measured two specific things:
- Entropy: How messy or chaotic the tumor's texture is. (High messiness = more chaos).
- Sphericity: How round the tumor is. (A perfect sphere is smooth; a jagged, lumpy tumor has low sphericity).
The Simulation: They used these two measurements to set the "speed limits" for their virtual tumor. If a tumor was very messy, the simulation made the virtual cells grow faster. If the tumor was very lumpy and irregular, the simulation made the center of the tumor more likely to die off (necrosis) because it might not have enough blood supply.
The Output: After running this virtual simulation, the computer generated six new numbers. These weren't just measurements; they were summaries of how the tumor might behave. They included things like "estimated growth rate" and "necrosis ratio."

The Super-Brain: TMAE

To make sense of the 3D CT scans, the researchers didn't just use a standard camera. They used a special type of AI called a Transformer-based Masked Autoencoder (TMAE). Imagine you show a computer a picture of a tumor but cover up 75% of it with black squares. The computer has to guess what's under the squares based on the visible parts. By doing this millions of times, the computer learns to understand the deep, 3D structure of the tumor without needing a human to label every single part. This "TMAE" became the main brain for reading the images.

Putting It All Together: The Four-Modality Fusion

The researchers then combined four different types of information into one giant prediction model:

The Simulation: The six new numbers from their virtual tumor game.
The Deep Brain: The complex patterns the TMAE found in the CT scan.
The Classic Stats: Hundreds of traditional measurements (radiomics) taken from the tumor.
The Patient Info: Basic facts like age, sex, and cancer stage.

They fed all this into a survival model to see who would live longer and who might not.

What They Found

The results were promising, but the authors are careful not to call it a magic cure.

The Score: On the Lung1 dataset, their new four-part model achieved a C-index of 0.641. In the world of survival prediction, a score of 0.5 is like flipping a coin, and 1.0 is perfect. The previous best method on this same dataset scored 0.631. So, they nudged the score up slightly.
The "What-If" Factor: The most interesting part was that the six simulation numbers (the "proxy features") added something new. Even though they were just six numbers, they helped the model perform better than using hundreds of traditional measurements alone. This suggests that capturing the interaction between the tumor's messiness and its shape (via the simulation) gives the computer a clue it was missing before.
The Best Case: When they tweaked the math behind the simulation a bit more (an "exploratory coefficient-optimization"), they saw the score go up to 0.662. However, the authors stress that this is an "exploratory" result, meaning it's a hint of potential rather than a final, locked-in victory.

What This Means (and What It Doesn't)

The paper suggests that adding a "virtual simulation" to the mix helps the computer understand the tumor better. It's like adding a weather forecast to a traffic report; you know not just where the cars are, but how the storm might change the traffic flow.

However, the authors are very honest about the limits:

It's a Simulation, Not Reality: The virtual tumor growth isn't a real measurement of how the patient's tumor actually grew. It's a mathematical guess based on the picture. They didn't have follow-up scans to prove the simulation was "right" about the biology.
One Test Only: They tested this on just one group of 390 patients. They haven't tried it on a different group of people yet, so we don't know if it works everywhere.
Not Ready for the Clinic: This is a research tool for now. It's not ready to be used by doctors to make treatment decisions today. It needs more testing, more data, and proof that it works on different types of scanners and hospitals.

In short, the paper shows that teaching a computer to play a "what-if" game with a tumor's picture can give it a slight edge in predicting the future. It's a small but meaningful step forward, suggesting that the secret to better predictions might lie in understanding how a tumor's shape and chaos dance together, rather than just counting them separately.

Technical Summary: Structured Proxy Features for Multimodal NSCLC Survival Prediction from Pretreatment CT

Problem Statement
Non-small cell lung cancer (NSCLC) remains a leading cause of cancer mortality, with survival stratification hindered by intratumoral heterogeneity that conventional descriptors fail to fully capture. Existing radiomic and deep learning approaches typically treat imaging features as independent quantities, overlooking structured biological interactions between tumor characteristics, such as the relationship between heterogeneity and proliferative capacity or boundary irregularity and necrosis susceptibility. While multimodal fusion of radiomic, deep, and clinical features has improved performance, gains on the public Lung1 benchmark have plateaued within a narrow C-index range of 0.577–0.631. This study investigates whether structured proxy features, derived from simulation of tumor dynamics, can provide complementary prognostic information to standard deep learning and radiomic representations.

Methodology
The proposed framework integrates four modalities: simulation-derived proxy features, Transformer-based Masked Autoencoder (TMAE) embeddings, conventional radiomics, and clinical covariates.

Radiomic-Parameterized Cellular Automaton (CA) Simulation:
- A rule-based cellular automaton simulates tumor growth dynamics on a 3D voxel lattice.
- The simulation is parameterized per-patient using two CT-derived radiomic descriptors: voxel intensity entropy ( $E_i$ ) to drive the proliferation coefficient ( $\alpha$ ) and shape sphericity ( $S_i$ ) to drive the necrosis coefficient ( $\beta$ ).
- The simulation models transitions between tissue states (normal, proliferating, viable tumor, necrotic, malignant subclone) governed by differential equations.
- From the simulation, six low-dimensional "proxy features" are extracted: the proliferation coefficient ( $\alpha$ ), necrosis coefficient ( $\beta$ ), entropy ( $E_i$ ), sphericity ( $S_i$ ), growth rate, and necrosis ratio. These features aim to summarize structured interactions between heterogeneity and morphology.
Imaging Backbone (TMAE):
- A Transformer-based Masked Autoencoder (TMAE) serves as the imaging backbone, selected after comparing it against a Convolutional Autoencoder (CAE) and a MedicalNet-ResNet18 (RN18).
- The TMAE is trained via self-supervised learning to reconstruct masked patches from 3D pretreatment CT tumor volumes ( $32^3$ voxels).
- It produces a 768-dimensional global embedding (CLS token) representing the tumor volume, utilizing self-attention to relate spatially distant regions.
Multimodal Fusion and Survival Modeling:
- The six simulation-derived features are concatenated with TMAE embeddings, selected radiomic features, and clinical variables (age, sex, stage).
- Two survival models are trained on this fused representation: a linear Cox Proportional-Hazards (CoxPH) model and a non-linear Neural Cox model (using a multilayer perceptron).
- An ensemble risk score is generated by averaging the min-max normalized outputs of both models to reduce sensitivity to the functional form of the risk-feature relationship.

Key Contributions

Structured Proxy Features: Introduction of a radiomic-parameterized cellular automaton that generates six compact, simulation-derived features to encode structured interactions between tumor heterogeneity and morphology.
Encoder Selection: A controlled comparison identifying the TMAE as the superior imaging backbone for this pipeline, outperforming CAE and RN18, likely due to its ability to capture distributed heterogeneity via self-attention.
Ablation and Sensitivity Analysis: Demonstration that the structured proxy features provide complementary value, with the simulation branch (SM) outperforming hundreds of selected radiomic features when paired with the TMAE. The study also includes sensitivity analyses confirming the numerical stability of the simulation parameters.

Results
Evaluated on the public Lung1 cohort (n = 390) using a fixed train/test split:

Primary Benchmark: The four-modality fusion (Simulation + TMAE + Radiomics + Clinical) achieved a C-index of 0.641 (iAUC 0.731, log-rank p < 0.001). This represents an improvement over the closest prior multimodal benchmark on Lung1 (C-index 0.631) [15].
Exploratory Optimization: An exploratory coefficient-optimization analysis, tuning the affine scaling coefficients for the simulation parameters, achieved a best observed C-index of 0.662 (iAUC 0.748).
Model Performance: The Neural Cox model consistently outperformed the linear CoxPH model, particularly when simulation-derived growth rate and necrosis ratio features were added, suggesting non-linear interactions in the data.
Feature Efficiency: The six-feature simulation branch (SM) combined with TMAE (TMAE+SM) outperformed the combination of TMAE and hundreds of radiomic features (TMAE+R), indicating that structured, low-dimensional summaries may capture prognostic structure lost in high-dimensional, independent radiomic sets.
Stability: Sensitivity analysis showed the pipeline is numerically stable to ±10% and ±20% perturbations in simulation parameters, with Spearman rank correlations remaining above 0.989.

Significance and Claims
The paper claims that structured proxy features derived from baseline CT descriptors can improve multimodal NSCLC survival prediction within a fixed benchmark framework. The primary significance lies in demonstrating that compact, rule-based summaries of tumor dynamics (growth and necrosis) provide complementary prognostic information beyond standard deep learning and radiomic descriptors.

The authors emphasize that the framework relies solely on routinely acquired pretreatment CT and standard clinical variables, avoiding the need for invasive assays or genomic profiling. The simulation-derived features are described as "auditable," as they trace directly to specific radiomic descriptors and defined transition rules.

Limitations and Scope
The authors maintain a modest stance regarding the clinical applicability of the results:

Single-Benchmark Observation: The results are based on a single public cohort (Lung1) with a fixed train/test split. Stability across alternative splits and external cohorts has not yet been established.
Proxy vs. Biological Validation: The simulation features are interpreted as data-driven prognostic proxies, not biologically validated reconstructions of patient-specific tumor growth. The affine coefficients were calibrated using survival data rather than longitudinal imaging or histopathology, which were unavailable in the Lung1 cohort.
Retrospective Ranking: The framework demonstrates retrospective relative risk ranking rather than calibrated probability estimation or threshold-validated risk stratification suitable for immediate clinical decision-making.
Deployment Requirements: The authors state that significant steps remain before deployment, including standardized tumor segmentation, calibration on local imaging protocols, and prospective multi-site validation.

In conclusion, the study establishes a foundation for future repeated-split and external-cohort studies, suggesting that integrating simulation-inspired descriptors with modern volumetric representations offers a practical approach for retrospective risk stratification.