[Paper Review] Exploring Explainable AI Techniques for Improved Interpretability in Lung and Colon Cancer Classification

Mukaffi Bin Moin, Fatema Tuj Johora Faria+3 more|arXiv (Cornell University)|May 7, 2024

Radiomics and Machine Learning in Medical Imaging4 citations

TL;DR

This study proposes a hybrid deep learning and explainable AI (XAI) framework for automated lung and colon cancer classification using histopathological images. It evaluates eight pre-trained CNNs (e.g., Xception, DenseNet201) with data augmentation and applies XAI techniques like Grad-CAM and SmoothGrad to enhance model interpretability, achieving up to 99.89% accuracy and superior visualization of decision-making processes.

ABSTRACT

Lung and colon cancer are serious worldwide health challenges that require early and precise identification to reduce mortality risks. However, diagnosis, which is mostly dependent on histopathologists' competence, presents difficulties and hazards when expertise is insufficient. While diagnostic methods like imaging and blood markers contribute to early detection, histopathology remains the gold standard, although time-consuming and vulnerable to inter-observer mistakes. Limited access to high-end technology further limits patients' ability to receive immediate medical care and diagnosis. Recent advances in deep learning have generated interest in its application to medical imaging analysis, specifically the use of histopathological images to diagnose lung and colon cancer. The goal of this investigation is to use and adapt existing pre-trained CNN-based models, such as Xception, DenseNet201, ResNet101, InceptionV3, DenseNet121, DenseNet169, ResNet152, and InceptionResNetV2, to enhance classification through better augmentation strategies. The results show tremendous progress, with all eight models reaching impressive accuracy ranging from 97% to 99%. Furthermore, attention visualization techniques such as GradCAM, GradCAM++, ScoreCAM, Faster Score-CAM, and LayerCAM, as well as Vanilla Saliency and SmoothGrad, are used to provide insights into the models' classification decisions, thereby improving interpretability and understanding of malignant and benign image classification.

Motivation & Objective

• Address the critical need for early and accurate diagnosis of lung and colon cancer, which remain leading causes of cancer-related mortality worldwide.
• Overcome limitations in histopathology-based diagnosis, such as inter-observer variability and reliance on expert pathologists, by automating classification using deep learning.
• Improve model interpretability in medical imaging by integrating multiple XAI techniques to visualize decision-making processes in cancer classification.
• Bridge the gap in explainable AI applications for lung and colon cancer classification using histopathological images, enhancing clinical trust and adoption.
• Evaluate and compare the performance of eight state-of-the-art pre-trained CNN architectures in a standardized setting to identify the most effective model for cancer detection.

Proposed method

• Employed eight pre-trained CNN architectures—Xception, DenseNet121, DenseNet169, DenseNet201, InceptionV3, ResNet101, ResNet152, and InceptionResNetV2—for transfer learning on histopathological image datasets.
• Applied advanced data augmentation strategies to improve model generalization and robustness across training and validation sets.
• Integrated multiple XAI techniques: Grad-CAM, Grad-CAM++, Score-CAM, Faster Score-CAM, LayerCAM for class activation maps, and Vanilla Saliency and SmoothGrad for saliency maps.
• Used class activation maps and saliency maps to visualize discriminative regions in histopathological images, highlighting features critical for malignant vs. benign classification.
• Evaluated model performance using standard metrics: accuracy, precision, recall, F1-score, Jaccard score, and log loss to quantify classification effectiveness.
• Hosted all code and trained models publicly on GitHub to ensure reproducibility and community access.

Figure 1: Representation of LC25000 dataset

Experimental results

Research questions

• RQ1 Which pre-trained CNN architecture achieves the highest classification accuracy for lung and colon cancer in histopathological images?
• RQ2 How do different explainable AI techniques (e.g., Grad-CAM, SmoothGrad) enhance the interpretability of deep learning models in medical image classification?
• RQ3 To what extent do XAI methods improve the visualization of discriminative regions in histopathological images for cancer diagnosis?
• RQ4 What is the performance trade-off between model accuracy and interpretability across different CNN architectures in this medical imaging task?
• RQ5 Can the integration of XAI techniques increase clinician trust and facilitate clinical decision support in automated cancer diagnosis?

Key findings

• Xception achieved the highest accuracy of 99.89% and the lowest log loss of 0.0384, outperforming all other models in the study.
• DenseNet201 achieved 99.71% accuracy with strong precision (99.71%), recall (99.71%), and F1-score (99.71%), indicating high consistency.
• InceptionResNetV2 performed the worst, with 97.65% accuracy and the highest log loss (0.8458), indicating lower confidence and higher error rates.
• Grad-CAM, Grad-CAM++, and SmoothGrad generated high-quality class activation and saliency maps that clearly highlighted malignant regions in tissue samples.
• The integration of XAI techniques enabled visualization of critical discriminative features in histopathological images, improving model transparency and clinical interpretability.
• Confusion matrices confirmed high specificity and sensitivity across all models, with Xception showing near-perfect class separation for both lung and colon cancer.

Figure 2: Representation of our proposed methodology for lung and colon cancer classification using explainable AI techniques