Melanoma is bad type of skin cancer. Today, a technique called dersmoscophy is used to detect melanoma. These method works with some features as ABC rule(Asymmetry, Border, Color)to decide whether lesion is melanoma or not. Now, automated detection is worked on due to early and easiness of detection. Main goal of this project is feature detection from skin lesion images.
Dataset features: Dataset includes 2000 train, 160 validation, 400 test images. Dimension of the images are 3843843(RGB). Additionally, superpixel and ground truth images are provided.
As a starting point, I work up skin lesion classification. For this purpose, I read the ISIC-2017 challenge paper and look at the algorithms of top ranked teams. I decided to use ResNet for classification task. For the following week I am planning to work on it. As an evaluation metric ROC AUC can be used for this part.
Secondly, ABC feature scores which are obtained from different annotators are exist in dataset. These features have information about whether lesion is melanoma or not. Aim of the project is improve the accuracy of classification by using these features. However, to obtain these features from annotators everytime is not efficient way. Therefore, they should be predicted and used as an input for classification. And these processes will work in parallel.
For introduction https://docs.google.com/document/d/1vpkeLOE9TGPD60vVkt1V31B5VhHm41phCnWM0klG4VA/edit?usp=sharing