Registration based fully optimized melanoma detection using deep forest technique
Abstract
Skin cancer could be considered as one of the critical forms of cancer, with Melanoma having a high mortality rate. If melanoma can be detected early on and thoroughly treated, the mortality caused by the disease may be prevented. Therefore, the Deep Forest Model a tree-based ensemble technique is used as a diagnostic tool for the detection of melanoma. The deep forest tree implemented contains the following feature extraction modules: SIFT, HoG, RIFT, LoG, and DoG. One of the important aspects of the deep forest models is Multi-Grained scanning which employs selectors to pick up elements like texture features. The selectors are built dynamically, allowing the training to end as soon as the ideal feature combination is identified. Image blur is corrected using kernel matrix estimation-based latent semantic analysis (KME-LSA), and then the mutual information-based image registration technique is used on medical images of melanoma. A cascade structure, which is introduced in the second module of Deep Forest, introduces level-by-level tree structures. To simplify the model, fewer hyper-parameters and a dynamic level-growing approach are implemented. The deep forest model's performance is contrasted with that of other classification models like SVM and Adaboost. Performance evaluation metrics such as Precision, recall, accuracy, f-measure, and time complexity analysis are used for assessing the classifier's performance. Due to the ensemble nature and feature extraction implementation in the deep forest model, the accuracy of the classifier is found to be 88.5 compared to SVM and AdaBoost models. Time complexity during testing of the model is calculated as 6.658 ms whereas other classifier models' performance is achieved in seconds. The performance analysis findings demonstrate that the Deep forest technique outperforms the current SVM and AdaBoost classifier with limited time complexity.