Active Deep Learning with Fisher Information for Patch-wise Semantic Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), excels in image segmentation but demands extensive labeled training data.
• Active Learning (AL) offers a solution by intelligently selecting minimal data for labeling, improving CNN performance efficiently.

Purpose of the Study:

• To propose a novel diversified Active Learning (AL) strategy based on Fisher Information (FI) for Convolutional Neural Networks (CNNs).
• To evaluate the efficacy of FI-based AL in brain extraction tasks for both semi-automatic segmentation of specific subjects and building generalizable universal models.

Main Methods:

• Developed a novel diversified AL strategy leveraging Fisher Information (FI) for CNNs.
• Utilized gradient computations from backpropagation for efficient FI calculation within the CNN parameter space.
• Evaluated the method on newborn and adolescent brain extraction under semi-automatic and universal model building scenarios.

Main Results:

• FI-based AL demonstrated improved performance with labeling less than 0.05% of voxels in both evaluated scenarios.
• The proposed FI-based AL significantly outperformed random sampling.
• Achieved higher accuracy than entropy-based querying in transfer learning for newborn brain extraction.

Conclusions:

• Fisher Information-based Active Learning is a highly effective strategy for improving CNN segmentation accuracy with minimal data.
• This method offers significant advantages over random sampling and entropy-based querying, particularly in transfer learning contexts.
• The proposed approach shows promise for efficient and accurate brain extraction in medical imaging applications.