Scholarlink

Conventional B-mode medical ultrasound and shear wave elastography (SWE) are essential tools for diagnosing and managing chronic liver disease. To enhance the effectiveness of real-time processing in AI systems for ultrasound image analysis, a pre-processing step to detect liver views is essential, as many abdominal ultrasound images do not include the liver. In this paper, we introduce a method for efficient liver view classification in ultrasound images using minimal labeled data and a simple neural network architecture. Our approach leverages active learning techniques to significantly reduce the labeling burden while maintaining high classification accuracy. We developed a straightforward convolutional neural network (CNN) named SimpleClassifier, which achieved outstanding performance compared to state-of-the-art models, such as ResNet-18 and MLP-Mixer, with fewer training samples. SimpleClassifier achieved an accuracy of 91.5% with 257 labeled frames, while ResNet-18 and MLP-Mixer required 801 labeled frames each to achieve 70.2% and 86% accuracy, respectively. These findings demonstrate that combining active learning with an AI classifier, regardless of its complexity, can improve the accuracy of the classifier with minimal labeled data. Code available at: https://github.com/abderhasan/SimpleClassifier .

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

Simple is More: Efficient Liver View Classification in Ultrasound Images Using Minimal Labeled Data and Simple Neural Network Architecture

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Abder-Rahman Ali,
  • Anthony E. Samir

摘要

Conventional B-mode medical ultrasound and shear wave elastography (SWE) are essential tools for diagnosing and managing chronic liver disease. To enhance the effectiveness of real-time processing in AI systems for ultrasound image analysis, a pre-processing step to detect liver views is essential, as many abdominal ultrasound images do not include the liver. In this paper, we introduce a method for efficient liver view classification in ultrasound images using minimal labeled data and a simple neural network architecture. Our approach leverages active learning techniques to significantly reduce the labeling burden while maintaining high classification accuracy. We developed a straightforward convolutional neural network (CNN) named SimpleClassifier, which achieved outstanding performance compared to state-of-the-art models, such as ResNet-18 and MLP-Mixer, with fewer training samples. SimpleClassifier achieved an accuracy of 91.5% with 257 labeled frames, while ResNet-18 and MLP-Mixer required 801 labeled frames each to achieve 70.2% and 86% accuracy, respectively. These findings demonstrate that combining active learning with an AI classifier, regardless of its complexity, can improve the accuracy of the classifier with minimal labeled data. Code available at: https://github.com/abderhasan/SimpleClassifier .