Abstract

This study is a proposal of an integrative framework using Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME) to provide trustworthy and explainable explanations in the context of a deep ensemble model designed to classify pulmonary carcinoma. Three different pre-trained models, ResNet50, DenseNet121, and EfficientNetB0, and a deep ensemble model were used in this research approach. The paper outlines the methodological pathway that starts with the preprocessing of the high-resolution computed tomography (CT) images, followed by image enhancement and feature extraction through convolution neural network architectures. As there is no clear suggestion of which pre-trained model gives the best performance as compared to others, an ensemble approach has been introduced. The ensemble approach is a combination of multiple deep learning models that combine predictions to provide better accuracy and less variance between the different model predictions. Two leading post-hoc explainability techniques, Grad-CAM and LIME, are used to make the model interpretable when applied to clinical situations. The experimental assessment performed on image datasets of lung cancer from the computed tomography (CT) modality demonstrates that ensemble learning with these XAI techniques increases the diagnostic outcomes and the transparency. The deep ensemble model achieved 0.972 of accuracy, while ResNet50, DenseNet121, and EfficientNetB0 achieved 0.94, 0.9358, and 0.94 of accuracy, respectively. The results clearly indicate the regions of interest and the diagnostic features that help in the detection of lung cancer and boost the clinical confidence in the system. Possible problems, including image quality differences and the computational load of ensemble methods, are also addressed, and methods of optimization and implementation are suggested. In conclusion, the results illustrate that the use of Grad-CAM and LIME in a deep ensemble framework is beneficial for achieving excellent prediction accuracy and delivering interpretable insights that are essential for clinical decision-making and follow-up treatment of patients.

Authors

Pragnya Das¹, Satya Narayan Tripathy², Kali Prasad Rath³
Centurion University of Technology and Management, India¹, Berhampur University, India², NIST University, India³

Keywords

Lung Cancer CT Imaging, ResNet50, DenseNet121, EfficientNetB0, Ensemble Learning, Grad-CAM, LIME, Model Interpretability