Roland-Tian / Medical-Skin-Lesion-Classifiers-Melanoma-Nevus-Seborrheic-Keratosis-Comparing-CNN-Architectures

Skin lesion classifiers for melanoma vs nevus vs seborrheic keratosis, comparing Squeeze & Excitation, SE-ResNeXt, Inception-ResNetV2 and Neural Architecture Search convolutional neural networks

Skin Lesion Classifiers for Melanoma, Nevus, or Seborrheic Keratosis comparing SENet-154, SE-ResNeXt-101, Inception-ResNetV2 and NASNet Convolutional Neural Networks with Transfer Learning (only last linear layers trained)

• Cross entropy loss, and Adam optimizer with default learning rate (0.001) used for all models.
• Later training showed promise with weight balancing and AdaBound optimizer: https://github.com/Luolc/AdaBound

Overall test set accuracy (random guessing would be expected to yield ~33% accuracy):

• SENet-154:
  64%
• SE-ResNeXt-101:
  61%
• Inception-ResNetV2:
  56%
• NASNet Large:
  57%

SENet-154 appeared to generalize the best.

SE-ResNext-101 reduced training error the most, but did not generalize as well.

Inception-ResNetV2 generalized the worst.

NASNet Large was the most resource intensive and took the longest to train.