Adaptive Fusion-Based Deep Ensemble Framework for Multi-Class Skin Cancer Classification Using Dermoscopic Images

Introduction: Skin cancer remains one of the most frequently diagnosed malignancies worldwide, with early diagnosis is vital for improving patient survival and allowing for timely treatment. Though deep learning-based computer-aided diagnosis systems have shown promising performance, many systems rely on single CNN or conventional ensemble methods with a fixed fusion techniques, thus limiting the potential for combining different architectures to take advantage of mutually beneficial information.

Methods: This study proposes an Adaptive Fusion-Based Deep Ensemble framework for multi-class skin lesion classification using dermoscopic images from the HAM10000 . Four state of art pretrained CNN, namely Xception, EfficientNetB4, InceptionResNetV2, and MobileNetV2 are fine-tuned using ImageNet pretrained weights, the Adam optimizer with a learning rate of 0.001 , and sparse categorical cross-entropy loss.The outputs of the fine-tuned models are then dynamically integrated through an Adaptive Fusion Module, which assigns instance-specific weights based on prediction confidence, entropy-based uncertainty, and model reliability.This adaptive fusion strategy enables the ensemble to assign optimal weights to the most reliable model(s) for each input image, thereby improving robustness and classification performance which is evaluated using standard classification metrics.

Results: This work presents a novel adaptive fusion framework for multi class skin cancer classification that achieved a test accuracy of 98.86%, validation accuracy of 98.70%, weighted precision of 98.88%, weighted recall of 98.86%, and weighted F1-score of 98.85%. Furthermore, the proposed ensemble outperformed the best individual model by 0.30% and achieved superior performance compared with recent ensemble-based approaches.

Conclusions: The proposed Adaptive Fusion-Based Deep Ensemble framework effectively integrates diverse feature representations learned by multiple pretrained CNN architectures through an instance-specific adaptive weighting mechanism. This enhances the robustness and reliability of multi-class skin lesion classification, demonstrating its potential for accurate computer-aided skin cancer diagnosis. Future work will focus on validating the framework on larger multi-center datasets and exploring its deployment in real-time clinical applications.