Skin cancer image classification optimization through transfer learning with Tensorflow and InceptionV3

Abstract

Skin cancer is a common illness that claims thousands of lives annually in the United States alone. Accurately identifying malignant tumors is crucial to survival but can be challenging as the visual distinctions between benign and life-threatening tumors are minimal. The purpose of this project is to explore deep learning algorithms that can be trained to systematically classify skin cancer images, create a program to execute the algorithm, and exemplify an optimization process for the program that can serve as a reference for future works. The project will adopt a transfer learning algorithm based on previous studies on the subject and select a pre-trained model given practical restraints. Then a program will be coded in Python to retrieve datasets, process images, train the model, and evaluate accuracy. Finally, the algorithm will be optimized by tuning model parameters and training restraints. The experiments revealed that the algorithm was able to perform the task with an accuracy of around 70%. Model parameters such as optimizer choice and learning rate and training restraints such as batch size and epoch count have significant impacts on the training results and require precise values for maxima accuracy and minimal overfitting.

Dermatological Decision Support Systems using CNN for Binary Classification

“…This study highlighted the effectiveness of deep learning-based approaches in increasing the accuracy and efficiency of dermatology diagnostics. In [12] deep learning algorithms were applied to classify skin cancer images. This study demonstrated an optimization method, involving transfer learning and pre-trained models.…”