GradFaith-CAM
Seeing Isn’t Always Believing: Evaluating Grad-CAM Faithfulness in Lung Cancer CT Classification
Author: Teerapong Panboonyuen • Accepted at 18th International Conference on Knowledge and Smart Technology (KST 2026)
🧠 Motivation
Grad-CAM has become the de facto explainability tool for medical image analysis.
But a critical question remains unanswered:
Do Grad-CAM heatmaps truly reflect the model’s reasoning — or are we just seeing convincing illusions?
This repository accompanies our KST-2026 accepted paper, providing a rigorous, quantitative evaluation of Grad-CAM faithfulness and localization reliability across modern deep learning architectures for lung cancer CT classification.
📄 Paper
Seeing Isn’t Always Believing: Analysis of Grad-CAM Faithfulness and Localization Reliability in Lung Cancer CT Classification
📍 KST 2026 (Accepted)
Authors:
- Teerapong Panboonyuen (Chulalongkorn University, MARSAIL)
🚀 Key Contributions
✅ Faithfulness-aware evaluation of Grad-CAM
✅ Cross-architecture analysis (CNNs vs Vision Transformers)
✅ Quantitative explanation metrics beyond visualization
✅ Exposure of shortcut learning and misleading saliency
✅ Clinical implications for trustworthy medical AI
🏥 Dataset
We evaluate on the publicly available IQ-OTH/NCCD Lung Cancer CT Dataset:
- 1,190 CT slices
- 110 patients
- Classes:
- Normal
- Benign
- Malignant
- Expert annotations by radiologists & oncologists
⚠️ All data are de-identified and ethically approved.
🧩 Models Evaluated
| Architecture | Type |
|---|---|
| ResNet-50 | CNN |
| ResNet-101 | CNN |
| DenseNet-161 | CNN |
| EfficientNet-B0 | CNN |
| ViT-Base-Patch16-224 | Transformer |
🔍 What Is GradFaith-CAM?
We go beyond pretty heatmaps.
✨ Faithfulness Metrics Introduced
1️⃣ Localization Accuracy
- Overlap between Grad-CAM maps and tumor regions
2️⃣ Perturbation-Based Faithfulness
- Drop in confidence when highlighted regions are removed
3️⃣ Explanation Consistency
- Stability across random seeds and model re-initializations
Together, these metrics answer a critical question:
Does the highlighted region actually matter for the prediction?
📊 Key Findings
🔥 Grad-CAM is NOT uniformly reliable
- CNNs often produce coarse or misleading attention
- DenseNet shows signs of shortcut learning
- ViT provides precise but sometimes non-faithful localization
- High accuracy ≠ trustworthy explanation
Seeing a heatmap does not mean believing the model.
🖼️ Visual Examples
⚙️ Installation
git clone https://github.com/yourusername/GradFaith-CAM.git
cd GradFaith-CAM
pip install -r requirements.txt
🧪 Run Experiments
Train a model
python experiments/train.py --config configs/resnet.yaml
Evaluate Grad-CAM faithfulness
python experiments/evaluate.py --model resnet50
Visualize explanations
python experiments/visualize.py --image sample.png
📌 Why This Matters
Medical AI does not fail loudly — it fails convincingly.
This work shows why blind trust in saliency maps is dangerous, and why explainability must be:
- Quantitative
- Model-aware
- Clinically grounded
📚 Citation
If you use this code, please cite:
@inproceedings{panboonyuen2026gradfaithcam,
title = {Seeing Isn’t Always Believing: Analysis of Grad-CAM Faithfulness and Localization Reliability in Lung Cancer CT Classification},
author = {Panboonyuen, Teerapong},
booktitle = {Proceedings of the 18th International Conference on Knowledge and Smart Technology (KST)},
year = {2026}
}
🤝 Acknowledgements
This research was conducted at Chulalongkorn University and MARSAIL (Motor AI Recognition Solution Artificial Intelligence Laboratory).
🧠 Final Thought
Interpretability without faithfulness is just another illusion.
Let’s build AI we can truly trust.