COMPARATIVE EVALUATION OF HAND-CRAFTED AND CONVOLUTIONAL FEATURES FOR MULTICLASS LUNG CANCER CLASSIFICATION FROM CT IMAGES

Abstract

The classification of lung cancer using computed tomography (CT) images is a significant problem in computer vision as the images in such cases have complex structural and textural patterns which may be difficult to describe consistently. In this study, two classical machine-learning pipelines and a small convolutional neural network (CNN) are comparatively evaluated using experiments for four class classification of lung CT images. A total of 988 PNG images were included in the dataset: 601 for training, 72 for validation, and 315 for independent testing, across four classes: adenocarcinoma, large-cell carcinoma, normal lung, and squamous-cell carcinoma. Histogram of Oriented Gradients (HOG) descriptors were fused with a linear support vector machine (SVM) and logistic regression, and a small CNN was trained on the fly using 64×64 gray images, class-weighted cross-entropy and horizontal flip image augmentation. HOG-logistic regression had the highest overall accuracy (0.616), macro F1-score (0.634) and macro ROC-AUC (0.845) on the independent test set, followed closely by HOG-linear SVM (0.610, 0.627, 0.820, respectively). The CNN achieved 0.422 accuracy, 0.430 macro F1-score, and 0.773 macro ROC-AUC. The confusion matrices indicated that malignant subtypes were significantly more challenging to identify as opposed to normal cases. The results show that a simple handmade gradient descriptor has the potential to deliver a good baseline result on a small CT-image dataset, while a small CNN that is trained from scratch needs further optimization or transfer learning to keep up with these classical approaches.