Related Experiment Video
Updated: Jan 11, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
729
Ensembling Vision Transformers and ResNet-50 for Interpretable Lung Cancer Diagnosis with Feature Fusion and XAI
Rahul1, Deborah Adedigba1, Raza Hasan2
1Department of Science and Engineering, Southampton Solent University, E Park Terrace, SouthamptonHampshire, SO14 0YN, UK.
Journal of Imaging Informatics in Medicine
|November 13, 2025
Summary
A new deep learning framework combining ResNet-50 and Vision Transformer (ViT) achieves over 99.8% accuracy in classifying lung cancer histopathology. This AI model offers a transparent and accurate solution, improving diagnostic reliability and patient outcomes.
Area of Science:
- Artificial Intelligence
- Computational Pathology
- Oncology
Background:
- Lung cancer diagnosis faces challenges due to inconsistencies and limitations in conventional methods, contributing to high mortality rates.
- There is a critical need for accurate, transparent, and clinically viable diagnostic systems in histopathological lung cancer classification.
Purpose of the Study:
- To propose and evaluate a novel deep learning framework for accurate and interpretable histopathological lung cancer classification.
- To address the limitations of current diagnostic methods by developing a robust AI-assisted system.
Main Methods:
- Developed a hybrid ensemble deep learning architecture integrating ResNet-50 for hierarchical feature extraction and Vision Transformer (ViT) for global context.
- Fused features from ResNet-50 (2048-D) and ViT (768-D) into a combined vector for classification.
- Employed a multi-disciplinary Explainable AI (XAI) strategy, including Grad-CAM, LIME, SHAP, and others, for model interpretability.
Main Results:
- The ensemble model achieved high accuracy: 99.96% (±0.0004%) mean cross-validation, 99.94% holdout test set, and 99.82% separate test set.
- Explainable AI methods demonstrated significant interpretability, with attention heatmaps showing 87.3% overlap with pathologist-identified regions of interest.
Conclusions:
- The proposed hybrid deep learning framework offers a highly accurate and interpretable solution for lung cancer histopathological classification.
- This AI-driven approach addresses current clinical gaps, showing potential to significantly improve diagnostic reliability and patient outcomes in lung cancer care.
