A method for lung cancer detection by means of explainable convolutional neural networks
- 1Institute of High Performance Computing and Networking (ICAR), National Research Council of Italy (CNR)
- 2Department of Engineering, University of Sannio
- 3Department of Medicine and Health Sciences “Vincenzo Tiberio”, University of Molise
Abstract
Lung adenocarcinoma currently represents the most prevalent subtype of lung cancer and a leading cause of cancer-related mortality worldwide. Accurate and early diagnosis is critical for improving patient outcomes, yet traditional histopathological assessment is time-consuming and dependent on pathologist expertise. We present a deep learning approach based on convolutional neural networks that achieves a 0.974 accuracy in the automatic classification of lung adenocarcinoma, squamous cell carcinoma, and normal tissue from histological images, which serves as the baseline for a novel quantitative framework evaluating model-attention heatmaps. To address the explainability challenges of deep learning in real-world clinical settings, we consider the adoption of several Class Activation Mapping techniques i.e., Gradient-weighted Class Activation Mapping, Gradient-weighted Class Activation Mapping++, Score-Weighted Class Activation Mapping, and Fast Score-Weighted Class Activation Mapping, to generate visual explanations behind the model’s predictions. To achieve this, we introduce a quantitative evaluation framework using Mean Squared Error, Root Mean Squared Error, Peak Signal-to-Noise Ratio, and Structural Similarity Index Measure. This framework confirms that the combination of Gradient-weighted Class Activation Mapping++ and Score-Weighted Class Activation Mapping produces the most consistent and structurally faithful heatmaps, with Structural Similarity Index Measure values of 0.94 for lung adenocarcinoma. By combining high diagnostic accuracy with qualitative visual insights, the proposed quantitative XAI validation framework aims to provide an explainable and consistent decision-support tool for pathologists, facilitating visual indication of model-attended regions contributing to the prediction within histological images.
Related
- Semantic edge detection of fractures in geological outcrops using edge aware dilated convolutional networks
- Explainable AI framework for skin lesion imaging based cancer detection
- Explainable Multimodal AI for In-Hospital Detection of Lung Cancer and Heart Failure Using Routine Data: Development, Internal Validation, and Simulated Federated Evaluation
- CBAM-enhanced convolutional networks for diabetic retinopathy classification in the ODIR dataset
- Attention-based Spatio-Temporal Graph Convolutional Networks-enhanced deep reinforcement learning for adaptive traffic signal control in urban traffic networks
BibTeX
@article{giammarco2026method,
title = {{A method for lung cancer detection by means of explainable convolutional neural networks}},
author = {Marcello Di Giammarco and Mario Cesarelli and Antonella Santone and Fabio Martinelli and Francesco Mercaldo},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-62967-5},
url = {https://doi.org/10.1038/s41598-026-62967-5},
note = {CC BY},
}
RIS
TY - JOUR
TI - A method for lung cancer detection by means of explainable convolutional neural networks
AU - Marcello Di Giammarco
AU - Mario Cesarelli
AU - Antonella Santone
AU - Fabio Martinelli
AU - Francesco Mercaldo
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-62967-5
UR - https://doi.org/10.1038/s41598-026-62967-5
AB - Lung adenocarcinoma currently represents the most prevalent subtype of lung cancer and a leading cause of cancer-related mortality worldwide. Accurate and early diagnosis is critical for improving patient outcomes, yet traditional histopathological assessment is time-consuming and dependent on pathologist expertise. We present a deep learning approach based on convolutional neural networks that achieves a 0.974 accuracy in the automatic classification of lung adenocarcinoma, squamous cell carcinoma, and normal tissue from histological images, which serves as the baseline for a novel quantitative framework evaluating model-attention heatmaps. To address the explainability challenges of deep learning in real-world clinical settings, we consider the adoption of several Class Activation Mapping techniques i.e., Gradient-weighted Class Activation Mapping, Gradient-weighted Class Activation Mapping++, Score-Weighted Class Activation Mapping, and Fast Score-Weighted Class Activation Mapping, to generate visual explanations behind the model’s predictions. To achieve this, we introduce a quantitative evaluation framework using Mean Squared Error, Root Mean Squared Error, Peak Signal-to-Noise Ratio, and Structural Similarity Index Measure. This framework confirms that the combination of Gradient-weighted Class Activation Mapping++ and Score-Weighted Class Activation Mapping produces the most consistent and structurally faithful heatmaps, with Structural Similarity Index Measure values of 0.94 for lung adenocarcinoma. By combining high diagnostic accuracy with qualitative visual insights, the proposed quantitative XAI validation framework aims to provide an explainable and consistent decision-support tool for pathologists, facilitating visual indication of model-attended regions contributing to the prediction within histological images.
SN - 2045-2322
ER -