Melanoma Detection CNN

Outline a brief description of your project.

Table of Contents

• General Information
• Conclusions
  • Initial Model - without any preprocessing
  • Model 2 - data augmentation
  • Model 3 - Dropout layer
  • Model 4 - Class rebalancing
• Tech Stack
• Acknowledgements
• Contact

General Information

To build a CNN based model which can accurately detect melanoma. Melanoma is a type of cancer that can be deadly if not detected early. It accounts for 75% of skin cancer deaths. A solution that can evaluate images and alert dermatologists about the presence of melanoma has the potential to reduce a lot of manual effort needed in diagnosis.

Conclusions

Following are the conclusions from different models tried and experimented on the dataset:

Initial Model - without any preprocessing

• Model is overfitting and this can be observed from
  1. the huge gap between accuracies of train and validation sets
  2. training loss going towards zero or nullifying but validation loss keeps increasing inversely
     
     

Model 2 - data augmentation

• Model is not overfitting but the accuracy has dropped
• model validation curve in both accuracy and loss plot is fluctuating
  
  

Model 3 - Dropout layer

• Model is not overfitting but the accuracy has dropped and compared to data augmentation model accuracy and loss curves are aligned
  
  

Model 4 - Class rebalancing

We can see train dataset is imbalanced and

• class seborrheic keratosis has lowest samples i.e., 58
• classes pigmented benign keratosis, melanoma, basal cell carcinom, nevus dominating the data in terms of proportion

• after rebalancing with augmentor package, the accuracy has been increased
• there is no overfitting happening
• train vs validation curves are smoother and not highly fluctuating like earlier models
  
  

Model	epochs	Comments
Initial Model	20	Model is overfitting
Model with Data Augmentation	20	Accuracy is low, overfitting is controlled
Model with Dropout	20	no overfitting, but plot curves are not smooth and fluctuating and less accuracy
Model with Class rebalance + Dropout	50	accuracy improved, no overfitting and curves are smoother

Tech Stack

• Python - v3.9.12
• tensorflow - v2.8.0
• keras - v2.8.0

Please refer requirements.txt file, for other optional libraries installed for this CNN exploration

Acknowledgements

• Project is developed as per requirements of assignment.
• Tensorflow - loading images and processing them
• Stackoverflow - Generating class distribution dictionary

Contact

Created by @Vinay26k - feel free to contact me!