Novel Transfer Learning Approach for Medical Imaging with Limited Labeled Data

Area of Science:

• Medical Image Analysis
• Deep Learning
• Artificial Intelligence

Background:

• Deep learning models require extensive data, which is scarce in medical imaging.
• Manual annotation of medical images is costly, time-consuming, and error-prone.
• Existing transfer learning methods using natural image datasets (e.g., ImageNet) are ineffective for medical images due to feature discrepancies.

Purpose of the Study:

• To propose a novel transfer learning approach for medical image analysis that leverages large unlabeled datasets.
• To develop a new deep convolutional neural network (DCNN) model incorporating recent advancements.
• To address the challenge of limited labeled data in medical image classification tasks.

Main Methods:

• Training a deep learning model on large unlabeled medical image datasets.
• Transferring learned knowledge to fine-tune the model on smaller labeled medical image datasets.
• Developing and evaluating a new DCNN architecture.

Main Results:

• The proposed transfer learning approach significantly improved classification performance in both skin and breast cancer scenarios.
• Skin cancer classification achieved an F1-score of 98.53% with the proposed method, compared to 89.09% when trained from scratch.
• Breast cancer classification accuracy reached 97.51% with the proposed method, versus 85.29% when trained from scratch. A diabetic foot ulcer classification task also showed significant improvement.

Conclusions:

• The proposed transfer learning method effectively overcomes data limitations in medical image analysis.
• The approach demonstrates potential for broad application in medical imaging tasks with abundant unlabeled data and limited labeled data.
• The developed DCNN model and transfer learning strategy enhance the performance of medical imaging tasks within the same domain.