Active Learning Framework for Medical Image Segmentation and Classification

Strategic Insights -

Research findings on Machine Learning - Computational Intelligence have shed light on a new active learning framework that optimizes the utilization of annotation resources in medical image segmentation and classification tasks. This framework, proposed by researchers at the College of Physics and Information Engineering, aims to reduce annotation costs by selecting more valuable samples for annotation from the pool of unlabeled data. The proposed framework is applicable to both 2D and 3D segmentation and classification tasks, and has been extensively validated on three publicly available and challenging medical image datasets.

Key Takeaways:

  • The proposed active learning framework is designed to optimize the utilization of annotation resources in medical image segmentation and classification tasks.
  • The framework relies on perturbation consistency evaluation to rank the consistency of each data sample, selecting samples with lower consistency as high-value candidates for annotation.
  • The framework has been extensively validated on three publicly available and challenging medical image datasets, including the Kvasir Dataset, COVID-19 Infection Segmentation Dataset, and BraTS2019 Dataset.
  • The experimental results demonstrate that the proposed framework can achieve significantly improved performance with fewer annotations in 2D classification and segmentation and 3D segmentation tasks.
  • The proposed framework is applicable to both 2D and 3D segmentation and classification tasks.
  • The researchers claim that the proposed framework enables more efficient utilization of annotation resources by annotating more representative samples, thus enhancing the model's robustness with fewer annotation costs.

Statistics:

  • 3 publicly available and challenging medical image datasets were used to validate the proposed framework: Kvasir Dataset, COVID-19 Infection Segmentation Dataset, and BraTS2019 Dataset.
  • 2D and 3D segmentation and classification tasks were evaluated using the proposed framework.
  • The experimental results showed a significant improvement in performance with fewer annotations in 2D classification and segmentation and 3D segmentation tasks.
  • The proposed framework was validated using perturbation consistency evaluation, which enables the selection of more valuable samples for annotation.

Sources:

  • "Pcdal: a Perturbation Consistency-driven Active Learning Approach for Medical Image Segmentation and Classification." IEEE Transactions on Emerging Topics in Computational Intelligence, 2025.
  • Tong Tong, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Tao Wang, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Xinlin Zhang, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Yuanbo Zhou, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Yuanbin Chen, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Longxuan Zhao, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.
  • Tao Tan, University of Fuzhou, College of Physics and Information Engineering, Fuzhou 350108, People's Republic of China.