Scholarlink

<p>Medical image diagnosis involves analyzing and interpreting medical images to identify potential health issues, and recent advances have increasingly incorporated machine learning and neural networks to improve diagnostic precision and computational efficiency. In computed tomography, volumetric scans are acquired to form three-dimensional representations of internal anatomical structures, where segmentation-based identification of critical regions is commonly employed to support diagnosis. However, conventional convolutional neural network pipelines indiscriminately process entire 3D volumes, including diagnostically irrelevant regions, leading to substantial computational overhead and reduced efficiency. To address this limitation, this study introduces a Channel Compressor Algorithm, designed to reduce redundant inter-slice information in three-dimensional computed tomography data while preserving diagnostically relevant features. The proposed method transforms stacked computed tomography slices into a compact multi-channel representation and applies channel reduction prior to convolutional neural network processing, enabling seamless integration with existing architectures. The effectiveness of the proposed approach is evaluated on an in-house dataset of approximately 260 patients for classifying the activity of Thyroid-associated orbitopathy. We demonstrate that the proposed channel compression consistently improves computational efficiency while maintaining or improving diagnostic performance across multiple convolutional neural network backbones. The method reduces model parameters by up to 92% and FLOPs by up to 97%, while achieving equal or superior classification performance. For instance, EfficientNetV2 with a compression ratio of 0.25 improves AUC from 0.870 to 0.960, and MobileNetV2 achieves an AUC of 0.927 with reduced computational cost compared to its uncompressed counterpart. This study shows that the channel compressor algorithm significantly improves efficiency for three-dimensional medical image analysis without compromising diagnostic accuracy and enables seamless integration with existing convolutional neural networks, providing versatility and ease of implementation in diagnostic applications.</p>

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

Three-dimensional convolutional neural network with channel compression module for efficient diagnosis of thyroid-associated orbitopathy

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • A-Seong Moon,
  • Jeong Kyu Lee,
  • Jaesung Lee

摘要

Medical image diagnosis involves analyzing and interpreting medical images to identify potential health issues, and recent advances have increasingly incorporated machine learning and neural networks to improve diagnostic precision and computational efficiency. In computed tomography, volumetric scans are acquired to form three-dimensional representations of internal anatomical structures, where segmentation-based identification of critical regions is commonly employed to support diagnosis. However, conventional convolutional neural network pipelines indiscriminately process entire 3D volumes, including diagnostically irrelevant regions, leading to substantial computational overhead and reduced efficiency. To address this limitation, this study introduces a Channel Compressor Algorithm, designed to reduce redundant inter-slice information in three-dimensional computed tomography data while preserving diagnostically relevant features. The proposed method transforms stacked computed tomography slices into a compact multi-channel representation and applies channel reduction prior to convolutional neural network processing, enabling seamless integration with existing architectures. The effectiveness of the proposed approach is evaluated on an in-house dataset of approximately 260 patients for classifying the activity of Thyroid-associated orbitopathy. We demonstrate that the proposed channel compression consistently improves computational efficiency while maintaining or improving diagnostic performance across multiple convolutional neural network backbones. The method reduces model parameters by up to 92% and FLOPs by up to 97%, while achieving equal or superior classification performance. For instance, EfficientNetV2 with a compression ratio of 0.25 improves AUC from 0.870 to 0.960, and MobileNetV2 achieves an AUC of 0.927 with reduced computational cost compared to its uncompressed counterpart. This study shows that the channel compressor algorithm significantly improves efficiency for three-dimensional medical image analysis without compromising diagnostic accuracy and enables seamless integration with existing convolutional neural networks, providing versatility and ease of implementation in diagnostic applications.