Detection of melanoma skin cancer using capsule network and multi-task learning frameworkdetection of melanoma skin cancer using capsule network and multi-task learning framework

Melanoma is the most dangerous and aggressive kind of skin cancer, which is also the most frequent form of cancer worldwide. Given the complexities involved, automatic melanoma detection using skin imaging has lately received interest within the machine learning field. Convolutional neural network has widely been employed in recent years to address this problem. However, existing CNN models for skin cancer classification have the drawback of ignoring crucial spatial relationship between features. They are only able to perform accurate classifications provided a predetermined set of features are present in the test data, regardless of how those features are distributed, which leads to false negatives. Furthermore, the CNN pooling layers responsible for down-sampling in these networks also result in loss of data and poor generalization performance. This study proposes a combination of convolutional block and Capsule Neural Network with a multi-task learning framework to address the aforementioned challenges and boost skin cancer classification. The model's efficiency was measured by a number of metrics, including accuracy, specificity, recall, and F1 score. The accuracy of the proposed model achieved 98.93%, 98.52%, 95.7%, and 98.87%, respectively, indicating great efficiency when compared to other existing networks. As a result, the proposed method offers less sophisticated and robust architecture for automating the process of melanoma diagnoses and accelerating detection procedures in order to save a life.