thirurjst / melanoma_classification

Melanoma classification using computer vision techniques on SIM-ISIC 2020 dataset

Melanoma_Classification

Repository to reproduce the results on SIM-ISIC Competition 2020 Dataset hosted on kaggle competitions

Steps performed:

1. Image Preprocessing:

   • Hair Artifact Removal using Bottom Hat Filter and inpainting.
   • Color Constancy corrections using Gray world and max RGB algorithms(Originally constructed by LincolnZjx)

2. Training CNN: Efficient Nets are trained in both ensemble and stand-alone manner

   • Stratified Group K-Fold Cross validation
   • Label Smoothing
   • Loss: FOCAL LOSS

3. Testing CNN:

   • Tested using both public and private test data hosted on kaggle competition

Metrics Used:

The same metric AUC which is given in kaggle is used to measure the performance of the model since the dataset is highly skewed.

PROGRESS

• Hair Artifact Removal
• Color Constancy
• Stratified Group K-Fold Split
• Label Smoothing
• Training
• Testing