Lung Cancer Detection with Vision Transformers

Overview

A deep-learning research project investigating Vision Transformers (ViT) for lung cancer detection from medical images. The goal was to evaluate whether attention-based architectures — which model global relationships across an image rather than purely local features — could be effective for medical image classification where subtle, spatially-distributed cues matter.

Approach

• Vision Transformer backbone — images are split into patches, embedded, and processed through self-attention layers, letting the model weigh relationships across the whole image instead of only local neighbourhoods the way a CNN does.
• Transfer learning — started from pretrained ViT weights and fine-tuned on the medical imaging dataset, a standard strategy for getting transformers to converge on the limited data typical of medical domains.
• Data pipeline — preprocessing and augmentation to expand effective dataset size and improve generalization, with a held-out split for honest evaluation.
• Implementation — built in a Jupyter / PyTorch workflow, keeping the training and evaluation loop transparent and reproducible.

What It Demonstrates

• Comfort with transformer architectures applied to vision, not just NLP.
• Practical handling of the realities of medical ML — small datasets, transfer learning, and careful evaluation.
• An experimental mindset: testing a then-emerging architecture against an established problem to understand its trade-offs.

An academic / research project focused on exploring the architecture and methodology rather than a deployed clinical system.