Scholarlink

Deep learning techniques have shown promise in fetal ultrasound image analysis, potentially saving clinicians time and providing better patient care. This study introduces a novel active learning model for fetal ultrasound image datasets, aimed at identifying valuable data for annotation to enhance model performance while reducing the need for extensive labeled data. Unlike traditional heuristic methods, this framework explicitly employs deep reinforcement learning to develop a data selection policy. Features from these images are extracted to represent the state space for the reinforcement learning model using a deep convolutional neural network. A deep Q-learning algorithm trains a Q-network, which outputs actions determining whether to annotate the data. Leveraging Deep Reinforcement Learning (DRL) to learn a data selection policy enables the model to achieve better performance with fewer labeled examples. The results showed 72.9% accuracy in fetal plane labeling within the active learning framework.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Deep RL-Based Active Learning Approach for Fetal Ultrasound Image Plane Labeling and Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Priya Pradiprao Phatale,
  • S. Gargi,
  • Simi Surendran

摘要

Deep learning techniques have shown promise in fetal ultrasound image analysis, potentially saving clinicians time and providing better patient care. This study introduces a novel active learning model for fetal ultrasound image datasets, aimed at identifying valuable data for annotation to enhance model performance while reducing the need for extensive labeled data. Unlike traditional heuristic methods, this framework explicitly employs deep reinforcement learning to develop a data selection policy. Features from these images are extracted to represent the state space for the reinforcement learning model using a deep convolutional neural network. A deep Q-learning algorithm trains a Q-network, which outputs actions determining whether to annotate the data. Leveraging Deep Reinforcement Learning (DRL) to learn a data selection policy enables the model to achieve better performance with fewer labeled examples. The results showed 72.9% accuracy in fetal plane labeling within the active learning framework.