Open Access

ARTICLE

Hybrid Attention-Enhanced Transfer Learning Framework with Particle Swarm Optimization for Explainable Skin Cancer Classification

Chanchal Biswas¹, Naveenkumar Mahamkali¹, Satish Anamalamudi¹, Bhaskar Marapelli², Ch Anil Carie^1,*

1 Department of Computer Science and Engineering, SRM University AP, Amaravati, Andhra Pradesh, India
2 Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, India

* Corresponding Author: Ch Anil Carie. Email:

Intelligent Automation & Soft Computing 2026, 41, 73-103. https://doi.org/10.32604/iasc.2026.087440

Received 16 June 2026; Accepted 28 August 2026; Issue published 21 September 2026

Abstract

Effective feature extraction is problematic due to subtle differences in skin texture, color, and shape. Moreover, imbalanced datasets are common in medical image analysis, which complicates classification by biasing models toward dominant classes, causing overfitting. The proposed approach is a hybrid one comprising transfer learning and self-attention. We have employed ResNet-50 pre-trained model as a feature extractor for dermatoscopic images. Max Pooling and Global Average Pooling were used to focus on the important patterns and reduce irrelevant background information. He normal kernel and L2-regularization initialization are applied to emphasize salient patterns and suppress irrelevant background information. The module of self-attention comes after the max-pooling and before the global average-pooling. This method allows for efficient modeling of lesion relationships across the entire globe in a dual pooling strategy. We have used particle swarm optimization (PSO) to tune the hyperparameters. Optimization methods, such as early stopping, ReduceLROnPlateau, dropout, and 10-fold cross-validation, help to guarantee strong convergence and thus avoid overfitting. In addition, we have used data augmentation and transfer learning to overcome the imbalance in the datasets. This will help to generalize across different lesion types. The proposed model achieves an accuracy of 97.77%, precision of 98.00%, recall of 97.17%, and F1-score of 97.45% on the HAM10000 dataset. To assess the robustness and generalization ability of the proposed solution, we compared using the ISIC-2019, ISIC-2019-2020, and ISIC datasets, which include both binary and multi-class classification. The proposed solution has also achieved better performance compared to several recent research works conducted to solve the challenges regarding skin cancer feature extraction and hence improve cancer detection accuracy. We have demonstrated that the proposed framework effectively focuses on clinically relevant lesion regions using Grad-CAM as an explainable AI (XAI) technique.

---

Keywords

---