Intelligent Labeling Based on Fisher Information for Medical Image Segmentation Using Deep Learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel in medical image segmentation but require extensive annotated data.
• Acquiring large medical datasets is challenging due to cost, time, and expertise.
• Active Learning (AL) strategies can reduce annotation burden by selecting informative samples.

Purpose of the Study:

• To propose and evaluate a novel active learning method utilizing Fisher Information (FI) for training CNNs in medical image segmentation.
• To demonstrate the efficacy of FI-based AL in improving model generalizability with limited target data.
• To assess the method's performance in brain extraction tasks across diverse target datasets.

Main Methods:

• Developed a new active learning method based on Fisher Information (FI) for CNNs.
• Employed efficient backpropagation and a novel low-dimensional FI approximation for computational feasibility in large CNNs.
• Evaluated the method on brain extraction using a patch-wise segmentation CNN in universal AL and semi-automatic segmentation scenarios.

Main Results:

• The proposed FI-based AL method significantly outperformed existing AL methods and baselines.
• The method demonstrated superior performance in improving model accuracy after annotating a very small fraction (<0.25%) of the target dataset.
• Effective performance was observed even when target datasets differed from source data in age group or pathology.

Conclusions:

• Fisher Information-based active learning is a highly effective strategy for training CNNs in medical image segmentation.
• This method substantially reduces the need for extensive data annotation, making CNN training more efficient.
• The FI-based AL approach offers a promising solution for developing robust medical image analysis models with limited labeled data.