Deep Learning Radiomics and Dosiomics Nomogram to Predict Radiation Pneumonitis Risk in NSCLC

Sci Rep 2025 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Predicting Lung Toxicity Before Starting Radiation Therapy

The Clinical Problem Radiation pneumonitis (RP) is a serious inflammatory lung injury that can occur weeks to months after radiotherapy for non-small cell lung cancer (NSCLC). Predicting which patients are at high risk before treatment allows clinicians to adjust radiation plans or monitor more closely.

Combined Model Innovation This study introduces the DLRDN (Deep Learning Radiomics and Dosiomics Nomogram), which integrates three complementary data sources: traditional radiomic features extracted from CT images, dosiomics features from radiation dose distributions, and deep learning features, all within a single predictive nomogram.

Multicenter Design The study enrolled 245 NSCLC patients from three hospitals, providing a multicenter validation framework essential for demonstrating that a prediction model generalizes beyond its training institution.

Published in Scientific Reports 2025 This work extends the field of radiation toxicity prediction by being among the first to systematically combine radiomic, dosiomics, and deep learning features in a single nomogram for RP in NSCLC.

TL;DR: The DLRDN model combines CT radiomics, radiation dosiomics, and deep learning features in a nomogram to predict radiation pneumonitis risk across 245 NSCLC patients from three hospitals.
Pages 2-3
Extracting Radiomic-Like Features from Radiation Dose Maps

What Is Dosiomics Dosiomics applies radiomic feature extraction methodology - originally designed for anatomical CT or MRI images - to the radiation dose distribution map generated during treatment planning. The dose map is a 3D volume describing how much radiation dose each voxel of tissue receives.

Why Dose Distribution Shape Matters Two patients can receive the same mean lung dose (MLD) but with very different spatial distributions. Dosiomics captures whether the dose is concentrated in a small region or spread across many small foci, whether it falls more centrally or peripherally, and whether there are high-dose hot spots in sensitive lung zones.

Complementary to Radiomics While CT radiomics captures the baseline anatomical state of the lung (fibrosis, emphysema, inflammation), dosiomics captures treatment-specific dose patterns. These two sources of information are largely independent, making their combination particularly powerful.

Feature Types Both first-order dose statistics (mean, maximum, percentile doses) and texture-based features of the dose distribution were extracted, analogous to the way GLCM and GLDM features are computed for image textures.

TL;DR: Dosiomics extracts radiomic-like texture and statistical features from the 3D radiation dose distribution, capturing spatial dose patterns that standard dosimetric parameters like mean lung dose miss.
Pages 3-4
Established Clinical Parameters That Predict Radiation Pneumonitis

V20 and V30 Lung Dose Volumes V20 is the percentage of total lung volume receiving at least 20 Gy; V30 is the percentage receiving at least 30 Gy. Both were identified as independent clinical predictors of RP risk, consistent with decades of clinical radiobiology research in lung radiotherapy.

Mean Lung Dose (MLD) MLD, the average radiation dose delivered to the entire lung volume, was also confirmed as an independent predictor. Higher MLD reflects greater total lung exposure and correlates with the probability of inflammatory injury.

Clinical vs. Imaging vs. Dose Features The study systematically compared the predictive value of clinical dosimetric parameters (V20, V30, MLD) versus CT radiomic features versus dosiomics features, finding that the combined DLRDN model outperformed any single source of information.

Thresholds in Practice Clinicians typically aim to keep V20 below 30-35% and MLD below 20 Gy in lung cancer radiotherapy. The DLRDN model refines this by accounting for individual patient lung characteristics that modulate sensitivity to these dose thresholds.

TL;DR: V20, V30, and mean lung dose were confirmed as independent RP predictors, but the combined DLRDN model incorporating radiomic and dosiomics features outperformed these clinical parameters alone.
Pages 4-5
Imaging Features That Capture Underlying Lung Vulnerability

DNNUN Feature DNNUN, a dosiomics non-uniformity texture feature, was positively correlated with RP risk. When the dose distribution is non-uniform (concentrated in patches rather than smoothly spread), the heterogeneous exposure pattern appears to increase inflammatory injury risk.

SRLGLE Feature SRLGLE (short run low gray level emphasis), a CT radiomic texture feature, was negatively correlated with RP risk. This feature is thought to reflect the underlying lung tissue texture, with certain baseline lung textures being more resilient to radiation-induced inflammation.

Deep Learning Features Ten deep learning features extracted from a CNN operating on the CT images complemented the seven traditional radiomic/dosiomics features, capturing complex non-linear patterns in lung parenchyma texture that hand-crafted features do not encode.

Feature Count After Selection From hundreds of candidate features, only 17 were ultimately selected for the DLRDN nomogram (7 radiomics/dosiomics features plus 10 deep learning features), striking the balance between predictive power and model parsimony.

TL;DR: The DLRDN incorporated 7 selected radiomics/dosiomics features plus 10 deep learning features; DNNUN non-uniformity predicted higher RP risk while SRLGLE texture indicated lung resilience.
Pages 5-6
Validation Across Three Hospitals

Training AUC The DLRDN achieved an AUC of 0.891 on the training set, indicating strong discriminative power for RP prediction at the patient level before treatment begins.

Internal Validation On an internal validation cohort from the same institution, the model maintained an AUC of 0.825, showing minimal overfitting and good generalization to held-out patients from the training site.

External Validation On patients from the two external hospital cohorts, the model achieved an AUC of 0.801 - a modest reduction from internal validation but still indicating clinically useful prediction performance across different treatment environments.

Calibration and Clinical Utility Beyond discrimination, the nomogram showed good calibration (predicted probabilities matching observed RP rates) and positive decision curve analysis, confirming that using the model for clinical decisions would provide net benefit compared to treating all patients as high or low risk uniformly.

TL;DR: The DLRDN achieved AUCs of 0.891 (training), 0.825 (internal validation), and 0.801 (external validation), with good calibration and positive decision curve analysis confirming clinical utility.
Pages 6-7
Toward Personalized Radiation Plan Adaptation

Adaptive Planning Application With a validated RP risk score before treatment, radiation oncologists could prospectively adjust radiation plans for high-risk patients - reducing V20, lowering MLD, or using proton therapy - before any injury occurs.

Expanding the Toxicity Spectrum Future work should extend the dosiomics-radiomics framework to other radiation toxicities in lung cancer treatment, including radiation esophagitis and cardiac toxicity, using similar feature extraction and integration methodologies.

Prospective Clinical Trial A randomized trial testing whether DLRDN-guided plan adaptation (versus standard planning) reduces RP incidence without compromising tumor control would provide the highest level of evidence for clinical adoption.

Immunotherapy Interactions As immunotherapy combined with radiotherapy becomes more common, RP risk is altered by immune activation. Future models should incorporate immunotherapy treatment status as a predictor variable alongside dosimetric and radiomic features.

TL;DR: The DLRDN model could enable prospective radiation plan adaptation for high-risk patients; future work should extend to other toxicities and account for the increasingly common use of concurrent immunotherapy.
Citation: Open Access, 2025. Available at: PMC12084522.