/ overview
A comparison of convolutional architectures for classifying dermatoscopic images of skin lesions, implemented in TensorFlow and Keras.
Transfer learning from ImageNet weights was combined with augmentation — flips, rotation, zoom and colour jitter — to counter a small and imbalanced dataset.
The best configuration reached 86% classification accuracy, with MobileNetV2 offering the strongest accuracy-to-size trade-off for a deployable model.
/ what it does
- 01InceptionV3, MobileNetV2 and VGG variants trained under identical conditions.
- 02Transfer learning plus fine-tuning of the upper convolutional blocks.
- 03Augmentation pipeline to reduce overfitting on a limited dataset.
- 04Class-weighted loss to handle imbalance across lesion types.
- 0586% accuracy on the held-out test split.
/ numbers
86%
accuracy
3 families
architectures
transfer learning
technique
/ stack
Modelling
TensorFlowKerasTransfer learning
Architectures
InceptionV3MobileNetV2VGG
Tooling
NumPyMatplotlibJupyter