Fusion of Features Extracted from Transfer Learning and Handcrafted Methods to Enhance Skin Cancer Classification Performance

Abstract

Nowadays, medical imaging has become crucial for detecting several diseases using machine learning and deep learning techniques. Skin cancer has been the most common of several diseases. If it has not treated early, a severe illness may cause in the patients, and this difficulty may lead to death. Several automated detection methods have been explored in the area, but the performance still does not reach the level needed by the medical sector. The common factors that affect the performance of the detection process are limited dataset, the method used for feature extraction, classification, hyperparameter tuning, and the like. This work proposed a fused feature extraction technique containing a transfer learning of the DenseNet-169 model and six handcrafted methods to capture richer and more detailed features. We applied a well-known machine learning algorithm called gradient boosting machine (GBM) for classification. We have used a publicly available dataset to train and evaluate our method: ISIC Archive datasets. The result shows that the proposed method improves the performance of skin cancer classification. GBM with the fused feature extraction technique is the highest performer with 87.91% of accuracy. Moreover, we use a few recent and best previously worked methods for comparison, using the same dataset. Our proposed method outperforms all of the previously worked algorithms by most of the evaluation metrics.