Multi-scale GC-T2: Automated region of interest assisted skin cancer detection using multi-scale graph convolution and tri-movement based attention mechanism

Melanoma skin cancer, an aggressive neoplasm arising from the malignant proliferation of melanocytes, represents a formidable challenge in the field of oncology due to its high metastatic potential and significant mortality rates. The timely identification of skin cancer plays a pivotal role in prevention and can substantially mitigate the incidence of certain skin malignancies, notably squamous cell carcinoma and melanoma, which tend to have a higher likelihood of successful treatment when detected in their early stages. Various researchers have suggested both automated and conventional methods for precise lesion segmentation to diagnose medical conditions associated with melanoma lesions. Nevertheless, the substantial visual resemblance among lesions and the significant intraclass variations pose challenges, resulting in the reduced accuracy in terms of performance. To alleviate these issues, we have proposed an automated skin cancer diagnosis framework known as Multi-scale GC-T2. In our work, we have utilized DermIS and DermQuest datasets in which several pre-processing techniques are applied in terms of Noise reduction and Data Augmentation using Median Enhanced Weiner Filter (MEWF) for enhancing the image quality. Besides, an Enriched Manta-Ray Optimization Algorithm (ENMAR) is adapted for ensuring the quality of pre-processed images. Also, for minimizing the model complexity, an appropriate lesion area is segmented accurately by the integration of semantic segmentation and DRL approach (i.e., Advanced Deep Q Network (AdDNet) and HAar-U-Net (HAUNT)). Following that, we designed a classifier Multi-scale GC-T2, where appropriate features are extracted using Multi-scale Graph Convolution Network (M-GCN). We have then proposed a punishment and reward mechanism for enhancing the feature processing, and tri-movement attention mechanism is utilized for minimizing feature dimensionality. Finally, the feature maps are fused using tri-level feature fusion module, and the sigmoid function is incorporated for classifying skin cancer. The proposed Multi-scale GC-T2 research is carried out using the MATLAB 2020A, and the performance of the proposed model is validated by evaluating metrics, such as accuracy, sensitivity, specificity and f1-score. The experimental results unequivocally highlight the proposed Multi-scale GC-T2 framework's superiority over existing models.