Skin Cancer Detection
Summary
Deep Learning is helping many industries across the world helping lots of people. One of the biggest impact is happening in health sector. We tried the latest CNN model's to detect benign and malignant mole which might turn out to be cancerous to the patient. First we wanted to detect the cancer on 7 types mole with 10015 images.
- Actinic Keratoses
- Basal cell carcinoma
- Benign keratosis
- Dermatofibroma
- Melanocytic
- Melanoma
- Vascular
The dataset: HAM10000
Data
There was a huge class imbalance in the dataset so we used tensorflow's image_generator dataset to create new images to have more balance in the data. Also, we are planning to create GAN model to create new images to get rid of the class imbalance
Model
We used ResNet50 to classify 7 types of moles with pre-trained weights which is provided by Keras. The Architecture
- 1 Normal Conv2d
- 1 Conv Block
- 2 Identity
- 1 Conv Block
- 3 Conv Block
- 1 Conv Block
- 5 Identity
- 1 Conv
- 2 Identity
- 1 Average Pool
- Flatten to Fully Connected layer
Results
| Attempt | Accuracy | Precision | Recall | #Notes |
|---|---|---|---|---|
| First | %76 | %70 | %72 | Baseline model |
| Second | %80 | %72 | %73 | ResNet50 |
| Third | %81 | %72 | %73 | ResNet50 + Image Generator |
Conclusions
The class imbalance issue was at hand and also it would be better to classify images as a binary classification task as benign or malignant . Which than let the doctors do their job. It is important at this task to not have Type I error where positive it malignant mole.