Case study 04 · 2023 · Medical imaging
SCED: Skin Cancer Early Diagnosis
Classifying skin lesions as benign or malignant with an EfficientNet that uses 92.6% fewer trainable parameters.
- Year
- 2023
- Field
- Medical imaging
- Stack
- EfficientNet, TensorFlow, Keras, scikit-learn
- Result
- 93.7%Test accuracy · published in IEEE
Overview
SCED classifies skin lesions as benign or malignant. I added involution layers to EfficientNet to make the model much lighter.
The EfficientNet-involution model reaches 93.7% test accuracy with 92.6% fewer trainable parameters and 9.5% less training time than the EfficientNet-B0 baseline, which scored 95.99%. Giving up a little accuracy for a far smaller model makes it practical to run in real time.
I also compared CNN, involution and EfficientNet models for accuracy and efficiency, and added an experimental severity score based on the model's predictions, lesion area and pigmentation.
Published at the 2023 10th International Conference on Computing for Sustainable Global Development (INDIACom), IEEE: “SCED: A Multi-criteria-based Modeling of Deep Learning Architectures for Skin Cancer Analysis.”