The objective of this project is to propose a methodology for the automatic classification of skin lesions, based on image analysis and machine learning.
The database is composed of 200 original images of pigmented skin lesions acquired by dermoscopy. 100 of them have been histopathologically confirmed as melanomas. Thus two classes are considered in this project: on the one hand confirmed melanomas and on the other hand the remaining benign lesions. In addition to the original images, binary images representing the region of interest of each lesion are included in the database. The segmentation of the images using superpixels are also added.
Morphological features are extracted from the segmented images thanks to the regionprops function from skimage.measure such as :
- area
- eccentricity
- perimeter
- equivalent diameter
- extent
- filled area
- minor axis length
- major axis length
- major axis length/minor axis length
- solidity
The most relevant features are then kept using ANOVA criterion
Texture features are extracted from the images thanks to the Local Binary Patterns algorithm The most relevant features are then kept using Pearson's correlation formula
The final dataframe is divided into a training set with 70% of the size of the original set and a test set with 30% of the size of the original set.
Multiple Machine Learning models are tested (Naive Bayes, SVM, Logistic Regression, Random Forest, Decision Trees, ...) and the performance of each model is assessed with measures such as :
- Accuracy
- F1 Score
- ROC Curve
- AUC
The max accuracy reached is 74% with Random Forests, which can be largely improved with state of the art models that are used today.