Optimized Faster Region Convolutional Neural Hybrid Kernel Multi-Birth Support Vector Machine for Pulmonary Disease Detection

Abstract

Accurate and early detection of pulmonary diseases is critical for controlling their spread and improving treatment outcomes. The standard molecular, antigen, and antibody tests are ineffective for early detection due to their longer testing duration and being expensive. Chest X-ray and Computed Tomography (CT)-based imaging methods can complement the standard techniques for early detection, but bias may easily affect them, resulting in lower sensitivity. Employing suitable Deep Learning (DL) methods with effective bias control can accurately identify the pulmonary disease patterns in the images within less time. This paper proposes an Optimized Faster Region-Convolutional Neural Hybrid Kernel Multi-Birth Support Vector Machine (OFRCN-HKMBSVM) classifier for improved pulmonary disease detection. With the minority class imbalance problem resolved by modifying the loss function with class weight adjustment, this model combines the Faster Region Convolutional Neural Networks (F-RCNN) and advanced Machine Learning (ML) model of Hybrid Kernel Multi-Birth Support Vector Machine (HKMBSVM) to detect the severity of pulmonary diseases without bias. This model extracts the region-based features from the pre-processed images using the F-RCNN utilizing the ResNet50 as the backbone model, whose last layer is replaced with HKMBSVM. This combination maps complex pulmonary disease patterns such as opacities, consolidations, or ground-glass patterns and identifies the linear and non-linear relationships between them to differentiate different pulmonary diseases. A hybrid kernel combining polynomial and Bessel kernel functions and the Red-billed Blue Magpie Optimizer (RBMO)--based hyperparameter tuning of the model helps achieve the objective. Evaluation results showed that the OFRCN-HKMBSVM-based model improved the early detection of pulmonary diseases with reduced bias and inference time.