An Ensemble of Transfer Learning Models for the Prediction of Skin Cancers with Conditional Generative Adversarial...

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision for Medical Diagnosis

Background:

• Early skin cancer detection is critical to prevent metastasis, but manual diagnosis is time-consuming, costly, and prone to errors due to visual similarities.
• Automated diagnostic systems are needed for improved multiclass skin cancer categorization.
• Deep learning models offer potential for accurate automated skin cancer classification.

Purpose of the Study:

• To develop and evaluate a fully automatic method for classifying multiple types of skin cancer.
• To investigate the impact of data augmentation techniques, including Generative Adversarial Networks (GANs), on the performance of deep learning models for skin cancer classification.
• To compare the performance of models trained on augmented versus unaugmented datasets and to assess the efficacy of an ensemble model.

Main Methods:

• Fine-tuning of deep learning models (VGG16, ResNet50, ResNet101) for skin cancer classification.
• Application of data augmentation using traditional image transformations and Conditional Generative Adversarial Networks (CGANs) to address class imbalance and enhance dataset realism.
• Development of an ensemble model combining predictions from multiple fine-tuned transfer learning models.

Main Results:

• Deep learning models (VGG16, ResNet50, ResNet101) achieved accuracies of 92%, 92%, and 92.25% respectively, after data augmentation.
• An ensemble of these models further improved classification accuracy to 93.5%.
• Models trained with appropriate data augmentation demonstrated superior performance compared to those trained on unbalanced datasets.

Conclusions:

• Data augmentation, particularly using GANs for realistic image generation, is crucial for improving the performance of deep learning models in skin cancer classification.
• The proposed automated method, especially the ensemble approach, shows significant potential for enhanced accuracy in multiclass skin cancer categorization.
• This approach offers a promising alternative to traditional diagnostic methods, potentially leading to earlier and more accurate detection.