Incorporating a Novel Dual Transfer Learning Approach for Medical Images.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Transfer learning in medical imaging reduces reliance on large labeled datasets.
• Existing methods face limitations due to domain mismatch between source and target data.
• Novel approaches are needed to enhance transfer learning performance in medical diagnostics.

Purpose of the Study:

• To propose Dual Transfer Learning (DTL), a novel approach to mitigate domain mismatch in medical image classification.
• To evaluate the effectiveness of DTL across multiple pre-trained models (VGG16, Xception, ResNet50, MobileNetV2).
• To assess DTL's performance on ISIC2020 (skin cancer) and ICIAR2018 (breast cancer) datasets.

Main Methods:

• DTL fine-tunes the last layers of pre-trained models using both unclassified and a small set of classified target task images.
• Data augmentation techniques are employed to balance classes and increase sample size.
• The approach is tested on VGG16, Xception, ResNet50, and MobileNetV2 models.

Main Results:

• DTL significantly improved the performance of all tested models, with improvements ranging from 0.28% to 34.76% with data augmentation.
• The Xception model demonstrated superior performance, achieving 96.83% accuracy for skin cancer and 99% for breast cancer classification.
• Significant performance gains were observed across all models, particularly when data augmentation was utilized.

Conclusions:

• Dual Transfer Learning (DTL) effectively enhances medical image classification by addressing domain mismatch.
• DTL improves the performance of various pre-trained deep learning models on cancer image datasets.
• The proposed method offers a promising solution for improving diagnostic accuracy in medical imaging with limited labeled data.