ERNet: A deep framework for detection and classification of lung cancer from histopathological images

Abstract

Lung cancer is a potent condition that impacts the mortality rate. The conventional assessment of lung cancer includes microscopic biopsies. The process comprises a labor-intensive visual assessment that is subjective and necessitates the expertise of a professional pathologist. However, noise typically affects the readable features in the low-resolution digital biopsy images. Moreover, the incidence of noise affects both inter- and intra-observer interpretations by pathologists. To overcome these issues, a novel enhanced RetinaNet (ERNet) has been proposed to detect and classify the lung abnormalities in a single framework. The proposed ERNet integrates a convolutional block attention module for the refined feature extraction. Additionally, the proposed ERNet employs a generalized intersection over union bounding box loss function to precisely localize abnormalities. The proposed method utilizes LC25000 lung histopathological images for its development. To improve, and denoise the lung biopsies image datasets, ant colony fourth-order partial differential equation has been applied. The comparative qualitative, and quantitative study has been presented with respect to existing methodologies such as faster regional convolutional neural network, single shot detector, RetinaNet and detection transformer. The quantitative assessments are evaluated in terms of accuracy, true positive rate, true negative rate, precision, F-score, Jaccard index, and Dice coefficient. The following values are obtained: 98.73%, 98.04%, 98.45%, 0.98, 0.98, 0.99, 0.98, 0.98, and 0.99, respectively. The results of qualitative, quantitative with ablation analysis exhibit that the proposed method surpasses the outcomes of the other pre-existing methods.