Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally

Area of Science:

• Biomedical Imaging
• Machine Learning
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) show great promise in biomedical image analysis.
• The widespread application of CNNs is hindered by the scarcity of large, annotated biomedical datasets.
• Manual annotation is time-consuming, expensive, and requires specialized expertise.

Purpose of the Study:

• To present a novel method, AIFT (active, incremental fine-tuning), to reduce the cost and effort of annotating biomedical images.
• To integrate active learning and transfer learning into a unified framework for efficient model training.
• To enhance CNN performance in biomedical imaging applications through cost-effective annotation strategies.

Main Methods:

• AIFT utilizes a pre-trained CNN to identify valuable, unannotated samples for expert annotation.
• The framework incorporates active learning to select the most informative samples.
• Incremental fine-tuning continuously updates the CNN with newly annotated data in each iteration.

Main Results:

• AIFT demonstrated a reduction in annotation costs by at least 50% across three diverse biomedical imaging applications.
• The method achieved significant performance improvements due to its active and incremental learning capabilities.
• The integrated approach effectively addresses the challenge of limited annotated data in the biomedical field.

Conclusions:

• AIFT offers a highly effective solution for reducing annotation costs in biomedical image analysis.
• The active, incremental fine-tuning approach enhances CNN performance while minimizing resource expenditure.
• This method facilitates the broader adoption of deep learning in biomedical imaging by overcoming data annotation barriers.