DSAL: Deeply Supervised Active Learning From Strong and Weak Labelers for Biomedical Image Segmentation.

Area of Science:

• Biomedical image processing
• Machine learning
• Computer vision

Background:

• Biomedical image segmentation is crucial for medical imaging analysis across modalities like microscopy and X-ray.
• Manual image annotation is costly, time-consuming, and labor-intensive, hindering the development of accurate AI models.
• Existing active learning methods often incur high computational costs and inefficiently use unlabeled data.

Purpose of the Study:

• To develop an efficient deep active semi-supervised learning framework (DSAL) for biomedical image segmentation.
• To reduce the annotation burden and computational costs associated with traditional active learning.
• To improve the utilization of unlabeled data and intermediate network knowledge.

Main Methods:

• Proposed a novel deep active semi-supervised learning framework (DSAL) integrating active and semi-supervised learning.
• Introduced a new criterion based on deep supervision to select informative samples for both strong and weak labelers.
• Leveraged disagreement in intermediate network features for efficient active sample selection and reduced computational load.

Main Results:

• DSAL effectively selects informative samples by considering both high and low uncertainty cases.
• The framework simultaneously generates oracle and pseudo labels in an ensemble learning approach.
• Experiments on multiple medical image datasets demonstrated DSAL's superior performance over state-of-the-art methods.

Conclusions:

• DSAL offers a computationally efficient and effective solution for biomedical image segmentation.
• The proposed method significantly reduces the need for extensive manual annotation.
• DSAL advances the application of active learning in the Internet-of-Medical-Things (IoMT) domain.