The applications of Artificial Intelligence (AI) in various fields are extensive and have the
potential to revolutionize various aspects of modern healthcare, for example, demonstrating
promising advances in improving the accuracy and efficiency of skin cancer detection and
classification. This area of study is of significant importance as it seeks to improve early
identification and diagnosis of skin cancer, positively impacting patient outcomes and treatment
strategies. This work describes a study carried out on the use of unbalanced datasets for the
classification of images of skin cancer lesions using Artificial Neural Networks, more specifically,
a dataset that has over 98% of samples belonging to the negative class. Three strategies were
applied to try to mitigate the difficulties caused by the large difference in the number of images
in each class, in this case, lesions that are melanoma and lesions that are not melanoma: reducing
the number of samples in the dataset to balance it, applying data augmentation and applying
class weights. In addition, methods for optimizing the training process of a Convolutional Neural
Network are successfully applied to automate the hyperparameters selection process and the
training time of models that use large neural networks as feature extractors is reduced because of
it. The data augmentation and class weights adopted in this work helped the training procedure
but were not enough to produce a large improvement in performance, but the latter method was
applied in the best result obtained.
|
