Semi advised learning and classification algorithm for partially labeled skin cancer data analysis

Development of automated diagnosis systems using machine learning and expert knowledge based data analysis requires effective automated learning models. However, models based on limited expert labeled training data can wrongly affect the results of diagnosis due to insufficient training knowledge acquired. On the other hand, getting more relevant analytical details from all the data used for training is an aspect that can enhance the efficiency of learning algorithms. This paper proposes a semi-advised training and classification algorithm that has the capability to effectively use limited labeled data along with abundant unlabeled data. It demonstrates the capability to use unlabeled data for training the algorithm by obtaining sufficient amount of information through incorporating an advised and/or partially supervised methodology. For comparative analysis, dermatological and histopathalogical images of skin cancer are used as experimental datasets. The proposed algorithm provided very impressive diagnosis outputs for both type of datasets in comparison to several other famous algorithms that are usually used in literature for classification.