Scholarlink

This research presents FreezeResNet, a novel deep-learning model designed to efficiently and accurately classify subtypes of intracranial hemorrhage. Building on the ResNet architecture, FreezeResNet enhances performance through selective layer freezing, tailored preprocessing, and fine-tuning techniques. Utilizing the extensive RSNA intracranial hemorrhage detection dataset, we addressed challenges related to dataset size and computational demands by eliminating unnecessary metadata and irrelevant images. We pre-trained and fine-tuned our model, iteratively adjusting parameters such as batch size, learning rate, and epochs to optimize performance. These refinements enabled FreezeResNet to achieve superior classification accuracy, surpassing established convolutional neural network (CNN) models including ResNet-50, InceptionNet, and AlexNet. Remarkably, the training is accelerated, resulting in rapid convergence to 91.29% accuracy after training for two epochs, improving to 99% with continued training. Despite the model complexity, FreezeResNet demonstrates significant speed improvement over traditional models, making it particularly suitable for clinical settings where timely and accurate intracranial hemorrhage detection is essential. Besides, we refined RSNA dataset and created a new dataset by removing metadata and irrelevant images. Project code and additional results are available at: https://github.com/SamaneSharifiMonfared/RSNA .

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

FreezeResNet: A Novel Deep Learning Model for Rapid and Accurate Classification of Intracranial Hemorrhage Subtypes

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Samane Sharifi Monfared,
  • Hassan Imani,
  • Lavdie Rada

摘要

This research presents FreezeResNet, a novel deep-learning model designed to efficiently and accurately classify subtypes of intracranial hemorrhage. Building on the ResNet architecture, FreezeResNet enhances performance through selective layer freezing, tailored preprocessing, and fine-tuning techniques. Utilizing the extensive RSNA intracranial hemorrhage detection dataset, we addressed challenges related to dataset size and computational demands by eliminating unnecessary metadata and irrelevant images. We pre-trained and fine-tuned our model, iteratively adjusting parameters such as batch size, learning rate, and epochs to optimize performance. These refinements enabled FreezeResNet to achieve superior classification accuracy, surpassing established convolutional neural network (CNN) models including ResNet-50, InceptionNet, and AlexNet. Remarkably, the training is accelerated, resulting in rapid convergence to 91.29% accuracy after training for two epochs, improving to 99% with continued training. Despite the model complexity, FreezeResNet demonstrates significant speed improvement over traditional models, making it particularly suitable for clinical settings where timely and accurate intracranial hemorrhage detection is essential. Besides, we refined RSNA dataset and created a new dataset by removing metadata and irrelevant images. Project code and additional results are available at: https://github.com/SamaneSharifiMonfared/RSNA .