Relative likelihood based aggregated dual deep neural network for skin lesion recognition in dermoscopy images

• Published in: Multimedia tools and applications Vol. 83; no. 21; pp. 60603 - 60626
• Main Authors: Anand, S., Sheeba, A., Maha Tharshini, M. K.
• Format: Journal Article
• Language: English
• Published: New York Springer US 2024; Springer Nature B.V
• Subjects: Ablation; Artificial neural networks; Classification; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Lesions; Machine learning; Multimedia Information Systems; Neural networks; Special Purpose and Application-Based Systems; Track 2: Medical Applications of Multimedia; Likelihood based add and compare decision; Phase congruency feature; Skin lesions classification; Gabor feature; Dermoscopy image; Dual deep neural network
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17908-z

This paper presents a novel approach classifying benign and malignant skin lesions based on dermoscopy images using a deep learning (DL) framework. We present a dual deep neural network (DDNN) to solve the problem of classification accuracy (ACC) that varies from class to class, which happens when different feature sets work better for different classes. This DDNN comprises two deep neural networks (DNN) designed to extract and characterize lesions, utilizing two distinct handcrafted feature sets: phase congruency features and (PC) Gabor features (GA). PC is known for its resilience to illumination and contrast variations, and GA captures high-frequency directional information in localized regions surrounding the lesion. The DDNN outputs of these two-trained networks are aggregated, employing a novel likelihood-based add-and-compare decision (LACD) in the output layer. With the utilization of training data, this proposed method generates additional discriminative features that aid in the classification of malignant and benign. To evaluate the effectiveness of our approach, ablation studies and experiments were conducted using skin cancer images from the ISIC archive, achieving a normalised ACC of 87.76% for ISIC2016 and 84.25% for ISIC2017. The performance of our method is also compared with existing skin lesion classification methods.