A Comparative Study of Vision Transformers and Convolutional Neural Networks for Lung Nodule Malignancy Classification in CT Imaging

Abstract

Accurate and timely malignancy categorization of pulmonary nodules in computed tomography (CT) images is critical for Health Information Technology, directly impacting clinical decision support systems and patient prognosis in lung cancer management and patient prognosis in lung cancer management. Although Convolutional Neural Networks (CNNs) are the standard, their local inductive bias can make them weak with regard to the modelling of the long-range, global contextual dependencies of medical images. While we recognize the natural restriction of evaluating 2D axial slices instead of full 3D volumetric data. This paper evaluates the effectiveness of a pre-trained self-supervised Vision Transformer (ViT) model to classify binary lung nodules, and leveraging the model's global self-attention mechanism to extract complex morphological features. Using a rigorously curated cohort of 2186 pulmonary nodule instances from the public LIDC-IDRI dataset, we preprocessed data via windowing, normalization, and resizing to 224×224 pixels. A ViT-Base model, pre-trained on ImageNet-21k, was fine-tuned and evaluated against a strong CNN baseline (DenseNet-121) using five-fold cross-validation. The ViT model achieved a superior F1-score of 0.891 (±0.018) and a mean AUC-ROC of 0.945 (±0.012) on the held-out test set. The results demonstrate that the Vision Transformer architecture presents a highly effective framework for this diagnostic task within HIT, surpassing traditional CNN-based approaches. Future work will focus on integrating 3D spatial information across multiple CT slices to further enhance model performance and clinical utility.