AN INTELLIGENT FRAMEWORK FOR LUNG CANCER DETECTION IN PET IMAGES USING ADVANCED HYBRID GCN-BiLSTM AND U-NET WITH ENHANCED DOVE SWARM OPTIMIZATION ALGORITHM (EDSOA)

Abstract

Patients who have lung cancer (LC) have a significant mortality rate from this fatal disease. Accurate LC staging and early diagnosis may preserve patients' lives. An estimated five million deaths occur each year as a result of LC. For both men and women, these mortality rates place it among the world's leading causes of death. Despite the adoption of various image processing (IP), biomarker-based, and machine automation techniques, healthcare providers still struggle to make an accurate and timely diagnosis of LC. This research aims to find malignant lung nodules (LN) on the input lung image and classify LC severity. An intelligent framework for LC detection (LCD) in PET images is offered in the suggested method. The suggested approach integrates modern techniques for feature extraction (FE), classification, segmentation, pre-processing, dataset collection, and hyperparameter tuning (HPT). In order to improve image quality and minimise noise, bilateral filtering (BF) is used for pre-processing. High-quality inputs for additional analysis are guaranteed by BF. Then, Regions of Interest (ROIs) are precisely segmented using U-Net for segmentation. Tumor boundaries are accurately delineated at the pixel level by U-Net. FE then obtains the features, including the Gray Level Run Length Matrix (GLRLM), Gray Level Co-Occurrence Matrix (GLCM), and the suggested Local Vector Pattern (LVP). For LC classification, a hybrid framework that combines attention mechanisms, Bidirectional (LSTM) Long Short-Term Memory (BiLSTM) units, and Graph Convolutional Networks (GCNs) is suggested. A comprehensive spatial and temporal FE are facilitated by this integration. The Enhanced Dove Swarm Optimization Algorithm (EDSOA) is introduced hyperparameter tuning for improve classification accuracy. Evaluated on the PET dataset, the framework EDSOA-HGNN-BiLSTM demonstrates robust accuracy in lung cancer detection, outperforming traditional methods.  Accuracy (A), Recall (R) (or Sensitivity (S)), Precision (P), and F-Score (or F-measure) (F1) were used to evaluate the classifiers. Positron Emission Tomography (PET)/Computed Tomography (CT) images collected from publicly available medical images (MI) of cancer from The Cancer Imaging Archive (TCIA) are used to evaluate the applicability of the suggested classifier and compare it with other current methods. According to the study of the outcomes, the suggested model executes better than the other methods currently in use in terms of A.