Malignant Melanoma Detection Using Ensemble Model and Improved BIRCH Clustering-Based Segmentation.

Abstract

Background: The most deadly disease in skin cancers family is melanoma. The color resemblance among melanoma-affected and healthy areas pose significant challenges in early detection.

Objective: An automated localization and segmentation of skin lesions at earlier stages remains challenging. To tackle these issues, a new method is proposed in this research for detecting malignant melanoma.

Method: This proposed strategy comprises five stages namely augmentation, preprocessing, segmentation, feature extraction, and classification. Initially, data augmentation is performed, then median filtering and image enhancement are applied to input image during preprocessing. Subsequently, IBIRCH algorithm is employed for segmentation. Furthermore, color and shape features, statistical features and improved local texton XOR pattern are extracted. Finally, ensemble model (proposed Bi-LSTM, CNN and DBN) is proposed which receives features and intermediate score obtained from each model undergoes improved score level fusion and yields final classification output.

Results: The proposed model is evaluated against traditional models and the suggested model achieved superior accuracy of 97.59% and 95.32% on datasets 1 and 2, respectively.

Conclusion: The improved performance of proposed model not only outperforms traditional approaches but also paves way for reliable automated early-stage melanoma diagnosis and so reduces life risk of patients due to this early detection.