Lung Cancer Classification from CT Images: A Comparative Study of SVM, Modified CNN, and EfficientNet–SVM Hybrid Models

Abstract This paper proposes a computer-aided diagnosis (CAD) framework that combines four complementary classification paradigms on the publicly available IQ-OTH/NCCD CT dataset with 1,097 slices across benign, malignant, and normal classes from 110 patients. The framework comprises: (i) an image-processing pipeline employing Gaussian smoothing, histogram equalization, Canny edge detection, and morphological closing, (ii) Otsu-based lung segmentation followed by handcrafted feature extraction combining statistical Local Binary Pattern (LBP), gradient, and shape descriptors, (iii) a Support Vector Machine (SVM) classifier benchmarked under a systematic ablation of three kernels and six class-balancing strategies (SMOTE, ADASYN, Borderline-SMOTE, Random Over-Sampling, SMOTE–Tomek, and cost-sensitive class weighting), (iv) a baseline convolutional neural network (CNN) and a modified CNN architecture enriched with batch normalization, and (v) a hybrid EfficientNet–SVM model that separates deep representation learning from maximum-margin decision-boundary fitting. The modified CNN achieves the highest overall test accuracy of 99.41% with a macro F1-score of 0.99, while the SVM with the RBF kernel under Random Over-Sampling reaches 98.81% balanced accuracy and the hybrid CNN–SVM attains 98.00% test accuracy. The result demonstrates both robust-ness across classes and competitiveness against recent state-of-the-art models on the same dataset.