Related Experiment Video
Updated: Jan 13, 2026
07:53
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
2.0K
Lung Nodule Malignancy Classification Integrating Deep and Radiomic Features in a Three-Way Attention-Based Fusion
Sadaf Khademi1, Shahin Heidarian2, Parnian Afshar1
1Concordia Institute for Information Systems Engineering, Montreal, QC H3G 1M8, Canada.
Journal of Imaging
|October 28, 2025
Summary
A new hybrid framework, I-VISTA, accurately classifies lung adenocarcinoma invasiveness using integrated visual, spatial, and temporal features. This deep learning and radiomic approach improves differentiation of early-stage from invasive lung cancers.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Lung adenocarcinomas presenting as subsolid nodules require accurate invasiveness assessment for optimal treatment.
- Differentiating between minimally invasive and invasive adenocarcinoma is critical for patient management.
- Current diagnostic methods can be limited in fully characterizing nodule invasiveness.
Purpose of the Study:
- To develop and evaluate a novel hybrid framework, I-VISTA, for assessing lung adenocarcinoma invasiveness.
- To integrate visual, spatial, and temporal features using deep learning and radiomic models.
- To improve the classification accuracy of subsolid nodules into different invasiveness groups.
Main Methods:
- A hybrid framework (I-VISTA) was developed, integrating three parallel processing paths: Shifted Window (SWin) Transformer for spatial features, Convolutional Auto-Encoder (CAE) Transformer for inter-slice relations, and a 3D Radiomic path for texture analysis.
- Features from these paths were fused using a Criss-Cross attention module for nodule classification.
- The framework was evaluated on 114 pathologically proven lung adenocarcinomas using a ten-fold cross-validation scheme.
Main Results:
- The I-VISTA framework achieved high performance metrics: 93.93% overall accuracy, 92.66% sensitivity, and 94.99% specificity.
- The Area Under the ROC Curve (AUC) was 0.93 ± 0.08.
- The hybrid approach integrating deep learning and radiomic models significantly outperformed standalone models in differentiating G1 (less invasive) from G2 (invasive) subsolid nodules.
Conclusions:
- The proposed I-VISTA framework demonstrates superior performance in classifying the invasiveness of lung adenocarcinomas presenting as subsolid nodules.
- Integrating multi-modal features through deep learning and radiomics offers a comprehensive approach for improved diagnostic accuracy.
- This hybrid model holds promise for enhancing clinical decision-making in lung nodule management.
