Chi2 weighted ensemble: A multi-layer ensemble approach for skin lesion classification using a novel framework - optimized RegNet synergy with Attention-Triplet

Abstract

Skin lesions, including various abnormalities and potentially fatal skin cancers, require early detection for effective treatment. However, current methods often struggle to identify the precise areas responsible for these abnormalities after model dominance dispersion. To address this, we propose a novel Transfer Learning-based framework that integrates Optimized RegNet Synergy architectures and Attention-Triplet mechanisms-comprising channel attention, squeeze-excitation attention, and soft attention-combined with an advanced Ensemble Learning strategy. A significant gap in current research is the lack of techniques for optimal weight allocation in model predictions. Our study fills this gap by introducing the [Formula: see text] Weighted Ensemble (CWE) method, which is further enhanced into a Multi-Layer [Formula: see text] Weighted Ensemble (ML-CWE) to improve model aggregation across multiple layers. Evaluation on the HAM1000 dataset demonstrates that our ML-CWE approach achieves an impressive accuracy of 94.08%, outperforming existing state-of-the-art methods. To enhance model interpretability, we employ Gradient Class Activation Maps (Grad-CAM) to highlight critical regions of interest, improving both transparency and reliability. This work not only boosts accuracy but also facilitates early diagnosis, addressing challenges related to time, accessibility, and cost in skin lesion detection, and offering valuable insights for practical applications in dermatology.

Fig 1. Sample images of each class.

Fig 2. Instance distribution for each class.

Fig 3. Schematic representation of methodology.

Fig 4. Images of the augmented samples.

Fig 5. ORNS architecture.

Fig 6. Feature extraction after activation of each layer (one sample).

Fig 7. Chi2 weighted ensemble in layer - L.

Fig 8. Organization of multi-layer CWE.

Fig 9. Confusion matrix obtained by RNXY architecture in CWE-Layer 3.

Fig 10. ROC-AUC curve obtained by RNXY architecture in CWE-Layer 3.

Fig 11. Confusion matrix obtained by RN_XY architecture in CWE-Layer 3.

Fig 12. ROC-AUC curve obtained by RN_XY architecture in CWE-Layer 3.

Fig 13. Confusion matrix obtained by RN architecture in CWE-Layer 4.

Fig 14. ROC-AUC curve obtained by RN architecture in CWE-Layer 4.

Fig 15. Step by step implementation of gradient class activation map.

Fig 16. GradCAM visualization for each class.

Fig 17. GradCAM visualization for AT explainability (Example by RNX002).