Analisis Komparasi Ekstraksi Fitur Manual dan Deep Learning untuk Klasifikasi Penyakit Paru Menggunakan SVM

This study performs a comparative analysis of three feature extraction scenarios: Handcrafted (LBP and GLCM), Deep Learning (VGG19), and Hybrid on a 4-class lung disease X-ray dataset. Images were processed at 224 x 224 pixels to preserve texture details. The Hybrid scenario utilized Global PCA for dimensionality reduction. Classification employed SVM with class weighting. Experimental results demonstrated that the standalone VGG19 achieved optimal performance with 96% accuracy, outperforming the Hybrid (91%) and Handcrafted (78%) scenarios. Feature analysis revealed that feature fusion triggered redundancy, where the dominance of handcrafted feature scores (Average 352.51) interfered with VGG19 features (Average 173.10), reducing majority class sensitivity. It is concluded that the single VGG19 architecture is more effective and efficient than the Hybrid method for this dataset.