Technical Summary: Digital Twin Modelling for Lung Cancer Management

Problem Statement

Lung cancer remains a leading cause of cancer-related mortality globally, necessitating early detection and personalized management to improve survival rates. Current methodologies face significant limitations, including a heavy dependence on extensive labelled datasets, susceptibility to overfitting, high computational demands, and a lack of model interpretability. Furthermore, conventional approaches often provide population-based survival estimates that fail to capture individual patient details, hindering the creation of dynamic, patient-specific virtual models capable of simulating tumour growth and predicting individual responses to therapies. There is a critical need for a system that integrates multimodal data (imaging and clinical metadata) to facilitate proactive, personalized healthcare decisions.

Methodology

The proposed research introduces a Digital Twin (DT) modelling engine integrated with deep learning-based survival analysis, risk stratification, and decision support. The framework operates through the following sequential modules:

1. Data Acquisition and Preprocessing

The system utilizes two primary data sources:

• Image Data: CT scan images from the IQ-OTH/NCCD-Lung Cancer Dataset, categorized into normal, benign, and malignant classes.
• Clinical Metadata: Patient attributes including age, gender, cancer stage, and molecular biomarkers from the "Lung Cancer Dataset."

2. Segmentation (TrEUNet++)

Input images are processed using a Trans-EfficientUnet++ (TrEUNet++) architecture. This model combines the EfficientNet backbone with Transformer modules within a Unet++ framework to extract multi-scale spatial features. It utilizes skip connections and transformer encoders to capture both local anatomical details and long-range global dependencies, producing precise segmentation of affected lung regions. From these segmented images, shape, size, and texture features are extracted (denoted as f1).

3. Survival Prediction and Risk Stratification (HC-WAE-SA)

The extracted image features (f1) and clinical metadata (f2) are fed into a Hybrid Convolution-based Weight Adaptive EfficientNetB7 with Spatial Attention (HC-WAE-SA) model.

• Architecture: This model employs hybrid convolutional layers (2D for images, 1D for text) followed by an EfficientNetB7 backbone. It integrates a Spatial Attention (SA) mechanism to dynamically weight critical regions within the medical images.
• Optimization: The Position Updated and Randomized Orangutan Optimization Algorithm (PUR-OOA) is employed to tune critical hyperparameters, including weight update values, learning rates, batch sizes, and activation functions. This algorithm enhances the balance between exploration and exploitation to minimize overfitting and ensure robust convergence.
• Outputs: The model predicts the expected survival time and classifies patients into low or high-risk categories.

4. Decision Support Recommendations (AGen-RLSTM)

Personalized treatment plans are generated using an Adaptive Generative Residual Long Short-Term Memory (AGen-RLSTM) model.

• Mechanism: This model integrates residual connections into LSTM networks to capture long-term temporal dependencies and employs a generative component to simulate future health states.
• Optimization: The same PUR-OOA algorithm is used to optimize the model's parameters for generating actionable treatment recommendations.

5. Feedback Loop and Incremental Learning

A clinician interface visualizes segmentation overlays, survival curves, and risk outcomes.

• Feedback Processing: Clinician feedback text is preprocessed and analyzed using a BERT-LSTM model to generate relevance scores.
• Incremental Learning: If feedback scores are low, the system triggers an incremental learning process. The HC-WAE-SA model is updated with new data without retraining from scratch, using PUR-OOA to optimize weight parameters, ensuring the Digital Twin evolves with clinical insights.

Key Contributions

The paper outlines six primary contributions:

1. Novel DT Framework: Development of a deep learning-based Digital Twin framework that simulates disease progression and personalizes treatment pathways using survival analysis and risk stratification.
2. Advanced Segmentation: Implementation of the TrEUNet++ model to accurately delineate tumour boundaries and distinguish subtle variations in shape and size, providing robust input for feature extraction.
3. Enhanced Survival Prediction: Introduction of the HC-WAE-SA model, which leverages hybrid convolutions and spatial attention to deliver highly accurate, patient-specific survival time estimations.
4. Optimized Parameter Tuning: Application of the PUR-OOA strategy to dynamically adjust model parameters, balancing exploration and exploitation to improve predictive accuracy and minimize overfitting.
5. Personalized Recommendations: Utilization of the AGen-RLSTM model to generate context-aware, personalized treatment plans that adapt to individual patient profiles.
6. Continuous Improvement: Establishment of a feedback-driven incremental learning loop using BERT-LSTM for score generation and PUR-OOA for weight optimization, ensuring the system adapts to evolving clinical insights.

Results and Evaluation

The proposed framework was evaluated against various benchmark models using standard metrics (Accuracy, Dice Coefficient, IoU, MAE, RMSE, MCC, etc.).

• Segmentation Performance: The TrEUNet++ model achieved an accuracy of 91.6% (WORST metric), outperforming Unet (0.847%), Unet++, FCN, and ResUnet. It also showed superior Dice Coefficient (0.915) and IoU (0.843) scores.
• Survival Prediction: The PUR-OOA-HC-WAE-SA model demonstrated significant improvements in convergence, achieving performance gains of 40% to 55% over other optimization algorithms (GOA, RBMO, WGOA, OOA) at the 10th iteration.
• Risk Stratification: The model achieved an accuracy of 94.667% (K-fold 3) and a specificity of 96.00%, outperforming literature models like CNN, ResNet, VGG-19, and EfficientNetB7.
• Recommendation Accuracy: In terms of Hit-rate and Mean Average Precision (MaP), the PUR-OOA-AGen-RLSTM configuration demonstrated superior capability in capturing relevant patterns compared to LSTM, RNN, GRU, and RLSTM variants.
• Statistical Robustness: The model exhibited lower Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) compared to existing DT-based models (LungDT, DT-LSTM, etc.) and standard deep learning architectures.

Significance and Claims

The authors claim that this research offers a transformative approach to lung cancer management by bridging the gap between data and clinical action. The significance of the work lies in its ability to:

• Personalize Care: Move beyond population-based estimates to provide patient-specific virtual twins that simulate biological and therapeutic dynamics.
• Enhance Decision Making: Support clinicians with evidence-based, real-time insights for diagnosis, risk stratification, and treatment planning, reducing reliance on trial-and-error mechanisms.
• Improve Outcomes: Facilitate early risk identification and timely interventions, potentially improving survival rates and resource utilization.
• Adaptability: Create a system that evolves through incremental learning and clinician feedback, maintaining clinical relevance over time.

The paper concludes that while the framework shows promising results in accuracy and robustness, future work should address limitations such as reliance on public datasets and the need for real-time integration with hospital systems to further validate clinical generalizability.