Analisa Kemampuan Deep Learning Untuk Deteksi Penyakit Paru Obstruktif Kronik (PPOK) Menggunakan Citra CXR Di RSP Ario Wirawan Salatiga

Abstract

Background: Chronic Obstructive Pulmonary Disease (COPD) represents a monumental global public health challenge, characterized by progressively worsening airflow limitation, high mortality, and substantial morbidity rates. Early and accurate detection plays a pivotal role in managing patient deterioration and improving quality of life. Traditional diagnosis in primary and secondary care heavily relies on expert interpretation of chest X-rays to detect subtle signs of hyperinflation and rule out comorbidities. However, this manual process is notoriously time-consuming, prone to inter-observer variability, and subjective. Consequently, this study aims to develop, optimize, and rigorously evaluate an automated computational detection system for COPD using a tailored Convolutional Neural Network (CNN) architecture based on standard digital X-ray imaging.

Methods: A quantitative, experimental computational approach was utilized with a dataset consisting of 276 chest radiograph images. The dataset was partitioned into 143 training images (100 Normal, 43 COPD) and 133 testing images (100 Normal, 33 COPD), deliberately maintaining a class imbalance to reflect real-world clinical prevalence. The CNN architecture was systematically evaluated across multiple hyperparameters, specifically training epochs and learning rates, to identify the absolute optimal model configuration for feature extraction and classification. Model performance was comprehensively measured using accuracy, sensitivity, and specificity metrics derived from confusion matrices.

Results: The empirical results demonstrated that the deep learning model achieved its highest testing accuracy of 98.5% at epoch 20 when paired with a learning rate of 0.1. At this optimal convergence state, the model demonstrated exceptional discriminatory power, yielding a sensitivity of 0.98 and a flawless specificity of 1 for the normal class. Conversely, for the critical COPD class, it achieved a sensitivity of 1 (zero false negatives) and a specificity of 0.98.

Conclusions: In conclusion, the implemented and optimized CNN architecture provides a highly accurate, robust, and rapid computational tool for COPD screening. With its perfect sensitivity for detecting pathological features, this system holds significant potential for integration as a clinical decision support system, particularly assisting clinicians in rural, high-volume, or under-resourced hospital environments.