An efficient ensemble deep learning model for lung cancer classification from histopathological images

Globally, lung cancer is the most commonly diagnosed cancer and the primary reason for cancer-related mortality. Early cancer detection increases the chances of successful treatments, thereby reducing patient mortality rates. Recently, medical imaging has been the most reliable technique to assist in diagnosing lung cancer and monitoring its treatment. Moreover, standard image classification tools often fall short in automatically identifying cancer, slowing down diagnoses. Therefore, automated systems for early lung cancer detection must be developed for greater sensitivity and accuracy. The emergence of deep learning techniques has enabled faster, more effective diagnosis and classification of lung nodules in the medical imaging field. Therefore, this article proposes an Ensemble Deep Learning Model for Lung Cancer Classification (EDLM-LCC) based on Medical Imaging. The proposed model aims to accurately classify lung cancer into adenocarcinoma, benign, and squamous cell carcinoma categories using advanced image analysis techniques. To accomplish this, the presented technique utilizes Non-Local Means filtering for image denoising. In addition, the EDLM-LCC system employs Convolutional Block Attention Module-ResNet50 as the feature extraction to obtain deep and discriminative feature representations from preprocessed images. For classification, an ensemble of Bidirectional Long Short-Term Memory, Temporal Convolutional Network, and Gated Recurrent Unit is adopted using a weighted majority voting strategy to improve robustness and classification accuracy. To enhance the performance of the presented method, the parameter selection process takes place using Bayesian optimization. The experimental evaluation of the presented EDLM-LCC system is carried out using a benchmark lung cancer histopathological images dataset, and the results indicate that the model demonstrates outstanding classification performance with an accuracy of 96.67% compared to existing techniques.