I've developed this project for the Melanoma Detection - IA Pierre Fabre challenge (Online). I've reached the 5th position with a VGG-16 and data preparation based on demartologic rules ABCDE.
This rule is used by demartologist and medecine student to detect melanoma on skin mole.
Melanoma lesions are often irregular, or not symmetrical, in shape. Benign moles are usually symmetrical.
Typically, non-cancerous moles have smooth, even borders. Melanoma lesions usually have irregular borders that are difficult to define.
The presence of more than one color (blue, black, brown, tan, etc.) or the uneven distribution of color can sometimes be a warning sign of melanoma. Benign moles are usually a single shade of brown or tan.
Melanoma lesions are often greater than 6 millimeters in diameter (approximately the size of a pencil eraser).
The evolution of your mole(s) has become the most important factor to consider when it comes to diagnosing a melanoma. Knowing what is normal for YOU could save your life. If a mole has gone through recent changes in color and/or size, bring it to the attention of a dermatologist immediately.
- Gaussian Blur Filter :
- Contrast agumentation :
- Crop :
-
class have been rebalanced with data augmentation (rotation, crop, color)
-
No transfer learning
- Vision based classification of skin cancer using deep learning, Simon Kalouche, Stanford
- Deep features to classify skin lessions
- Fully Convolutional Networks to Detect Clinical Dermoscopic Features *Knowledge Transfer for Melanoma Screening with Deep Learning
- Skin Lesion Analysis towards Melanoma Detection Using Deep Learning Network
- Melanoma detection using deep learning technology
- An Overview of Melanoma Detection in Dermoscopy Images Using Image Processing and Machine Learning