Automated Lung Cancer Classification from CT Scans using Deep Learning and Knowledge Distillation

Lung cancer remains a leading cause of cancer-related mortality worldwide, primarily due to late-stage diagnosis. Early and accurate detection from Computed Tomography (CT) scans is critical but is often hindered by the time-consuming and subjective nature of manual interpretation. This paper presents a deep learning-based system for the automated classification of lung CT images into Normal, Benign, and Malignant categories. We leverage an MLOps-driven pipeline to ensure reproducibility and scalability. The system employs transfer learning with VGG16 and Vision Transformer (ViT) architectures and introduces knowledge distillation to train a lightweight, efficient student model (VGG16) from a high-performing teacher model (ViT). Our dataset of 2,097 CT images was balanced using data augmentation and oversampling. The final distilled VGG16 model achieves a test accuracy of 98.09%, matching the performance of the larger teacher model while being significantly more efficient. This demonstrates that knowledge distillation can produce clinically viable models that balance high accuracy with the practical requirements for deployment in real-world healthcare settings.