This project classifies 8 different types of skin lesions using efficient lightweight mobile architectures. In all baselines, we managed to achieve at least 87% validation accuracy.
We chose various deep learning models that are strong contenders for being lightweight while also giving high accuracies.
We decided to choose a difficult dataset known as ISIC-2019. It provides various skin lesion categories and contains heavy class imbalance. Here are examples of the different categories.
This image below shows the various statistics of the ISIC-2019 training set.
Here we show the baseline results for FasterNet as it's the best model with the highest reported validation accuracy of 89.3%.
Here are the training and validation accuracies for the different epochs.
From the confusion matrix, it's clear that some classes are difficult to distinguish due to high inter-class variation.
We also investigate the tsne plot to ensure that embeddings are well separated near the classification layer. It also shows us which categories are similar to each other.
- Experimenting with various late feature fusion strategies.
- Then experiment with hierarchical bilinear pooling fusion (HBP).
- ISIC-2019 authors for their provided dataset.
- PyTorch Image Models for the various pretrained models on ImageNet.