Abstract:
Deep Learning has been generally used in the last few years to increase diagnosis accuracy in
medical imaging tasks including classification and segmentation. Advanced deep learning
algorithms use layers of neural networks to extract key visual properties, essential image
features, making the learning process increasingly difficult as more layers are added.
Comparing deep learning to traditional machine learning methods reveals a number of
advantages, but it is dependent on huge, annotated datasets, which are frequently constrained
by publicly accessible resources. This thesis suggests a practical method for overcoming this data
shortage that improves the precision of skin lesion diagnosis, particularly for the classification of
melanoma, by combining a collection of convolutional neural networks (CNNs) with generative
adversarial network (GAN) based augmentation.
With a small dataset, the first suggested model performed with an accuracy of 93.36%. The
second suggested method used conventional augmentation to increase the dataset, and it
provided superior data validation and accuracy of 60.16%. The training dataset was increased
using the final proposed method using both conventional and GAN based augmentation
techniques. With greater data validation, it performed better, with a performance accuracy of
95.01%. Additionally, it avoided trapping into local maxima. Which improved performance. When
dealing with the little training dataset, the scheme using the methods could be a preferable
