Authors: BAIDAA AL-BANDER, QAHTAN M. YAS, HUSSAIN MAHDI, RWAYDA KH. S. AL-HAMD
Abstract: Skin cancer is one of the most common cancers worldwide caused by excessive development of skin cells. Considering the rapid growth of the use of deep learning algorithms for skin cancer detection, selecting the optimal algorithm has become crucial to determining the efficiency of computer-aided diagnosis (CAD) systems developed for the healthcare sector. However, a sufficient number of criteria and parameters must be considered when selecting an ideal deep learning algorithm. A generally accepted method for benchmarking deep learning models for skin cancer classification is unavailable in the current literature. This paper presents a multi-criteria decision-making framework for evaluating and benchmarking deep learning models for skin cancer detection based on hybridisation of entropy and VIseKriterijumska Optimizacija I Kompromisno Resenje (VIKOR) methods. Twelve well-known convolution networks are evaluated and tested on eleven publicly available image datasets to achieve the target of the study. Several criteria related to deep convolutional neural networks (CNNs) architectures, including optimisation technique, transfer learning, class balancing, transfer learning, data augmentation, and network complexity, have been considered in the multi-criteria evaluation. The decision matrix (DM) is designed based on a crossover of the five evaluation criteria and twelve (CNNs) classification models on different datasets. Subsequently, in the benchmarking and ranking of deep learning classification models, multi-criteria decision making (MCDM) techniques are used. The MCDM uses a scheme that involves the integration of entropy with VIKOR approaches. For the weight calculations of evaluation criteria, entropy is applied, while VIKOR is used to benchmark and rank the models. The obtained results reveal that the InceptionResNetV2 model gained the first rank and is selected as the optimal architecture for skin cancer detection considering the five criteria investigated in our study. The presented framework achieves a significant performance in selecting the best algorithm, which could provide substantial guidance to the researcher working in the field.
Keywords: Skin cancer, deep convolutional neural networks, benchmarking
Full Text: PDF