A hybrid TriMDConv-UNet and EOO-based KAM-SANet framework for skin cancer identification
Jothi Prabha Appadurai1, Nirmala M2, Manjunatha B3 and Arunadevi Thirumalraj4
Professor, Department of Computer Science and Engineering,New Horizon College of Engineering, Ring Road, Bellandur Post, Bengaluru- 560103,Telangana,India2
Principal, New Horizon College of Engineering,New Horizon College of Engineering, Ring Road, Bellandur Post, Bengaluru- 560103,Telangana,India3
Research Scholar, Department of Computer Science and Engineering,Karunya Institute of Technology and Science, Coimbatore – 641114,Tamil Nadu,India4
Corresponding Author : Jothi Prabha Appadurai
Recieved : 04-May-2024; Revised : 26-August-2025; Accepted : 01-September-2025
Abstract
Skin cancer, a highly prevalent and potentially fatal disease, necessitates accurate and reliable diagnostic tools. This study presents a comprehensive methodology for skin cancer analysis using the International Skin Imaging Collaboration (ISIC)-2017 and ISIC-2019 datasets, through advanced techniques applied across the entire pipeline, from image preprocessing to classification. In the preprocessing phase, a guided filter is employed to reduce image noise, thereby minimizing unwanted artifacts and improving overall image quality. A novel network model, triple modified deep convolutional-UNet (TriMDConv-UNet), is introduced, offering three key innovations over the traditional UNet: (1) modified node connectivity, (2) the use of dilated convolutions instead of standard convolutions, and (3) the incorporation of multi-scale input features along with dense skip connections (DSC) in place of conventional skip connections (SC). For feature extraction, hierarchical relationships are captured using a capsule neural network (CapsNet), which provides more robust and discriminative feature representations. Additionally, a new self-attention network (SANet) based on the ResNet50 architecture, termed kernel attention mechanism (KAM)-SANet is proposed. This model integrates a KAM to enhance classification accuracy. Attention modules are utilized to extract contextual dependencies from images, emphasizing critical features essential for skin cancer detection. In the classification stage, the Eurasian oystercatcher optimiser (EOO) is employed to fine-tune hyperparameters. This nature-inspired optimization algorithm adjusts model parameters to achieve optimal performance in differentiating between benign and malignant skin lesions. The proposed model demonstrates superior performance in both segmentation and classification tasks. For the ISIC-2017 dataset, segmentation accuracy reached 95.03%, while for the ISIC-2019 dataset it was 96.20%. In classification, the model achieved 99.35% accuracy on ISIC-2017 and 99.65% on ISIC-2019. These results indicate that the proposed approach significantly outperforms existing models, delivering improved outcomes in skin cancer analysis.
Keywords
Skin cancer detection, Medical image segmentation, TriMDConv-UNet, Capsule neural network (CapsNet), KAM-SANet, Eurasian oystercatcher optimiser (EOO).
Cite this article
Appadurai JP, M N, B M, Thirumalraj A. A hybrid TriMDConv-UNet and EOO-based KAM-SANet framework for skin cancer identification. International Journal of Advanced Technology and Engineering Exploration. 2025;12(130):1414-1431. DOI : 10.19101/IJATEE.2024.111100699