Skin Cancer Detection (Kaggle Competition)
This project involved training an 2D image classifier for skin cancer detection with (optional) meta-features for a Kaggle Competition.
I fine-tuned the EfficientNet B1 model [Tan & Le, 2020] with a new classifier head to match the binary prediction of this task. I add a simple shallow network for passing meta-features to the network if the user desires, which will be concatenated with the EfficientNet output features before classification. I rely on PyTorch for modelling and Lightning for training. I fine-tuned the model locally using an NVIDIA TITAN Xp GPU.
I obtained a public LB score of 0.137. The score represents a fractional AUC with minimum score of 0 and maximum score of 0.2.