~/projects/onco
All projectsshippedMedical Imaging · Deep Learning2024
Lung Cancer Classification (Deep CNNs)
Trained and fine-tuned five CNN architectures (VGG19, MobileNet, ResNet, EfficientNet, Inception) for lung-cancer image classification.
problem
The problem
Medical image classification lives or dies on generalization, a model that memorizes the training set is worse than useless in a clinical context. The question isn't 'can one model fit' but 'which architecture generalizes, and why'.
approach
How I approached it
- Trained and fine-tuned VGG19, MobileNet, ResNet, EfficientNet and Inception on the same image dataset for a like-for-like comparison.
- Applied data augmentation and class balancing to fight overfitting and skew.
- Ran result analysis across architectures to reason about generalization.
architecture
Architecture
1Prepare
Augmentation + class balancing
2Transfer
Fine-tune 5 pretrained backbones
3Evaluate
Compare generalization across architectures
4Analyze
Error analysis → which backbone, and why
decisions
Engineering decisions & tradeoffs
$
Five architectures, not one
// Comparative evaluation is the actual research contribution, it tells you which inductive biases suit the data.
$
Augmentation + balancing first
// On medical datasets, data hygiene moves the metric more than model choice does.
stack
PythonTensorFlowKerasOpenCVNumPy
what's next
- › Grad-CAM overlays so predictions are explainable to clinicians.
- › Ensemble the top backbones for a small accuracy bump.