通过自蒸馏混合注意网络增强儿童头部MRI年龄估计

IF 4.9 2区医学 Q1 ENGINEERING, BIOMEDICAL

Biomedical Signal Processing and Control Pub Date : 2025-10-09 DOI:10.1016/j.bspc.2025.108748

Lei Shi , Xinran Huang , Wenchi Ke , Hrvoje Brkić , Yuchi Zhou , Ting Lu , Xian’e Tang , Lirong Qiu , Shuai Luo , Xingtao Zhang , Ziqi Cheng , Yushan Lin , Peixi Liao , Hu Chen , Yi Zhang , Yijiu Chen , Zhenhua Deng , Fei Fan

{"title":"通过自蒸馏混合注意网络增强儿童头部MRI年龄估计","authors":"Lei Shi , Xinran Huang , Wenchi Ke , Hrvoje Brkić , Yuchi Zhou , Ting Lu , Xian’e Tang , Lirong Qiu , Shuai Luo , Xingtao Zhang , Ziqi Cheng , Yushan Lin , Peixi Liao , Hu Chen , Yi Zhang , Yijiu Chen , Zhenhua Deng , Fei Fan","doi":"10.1016/j.bspc.2025.108748","DOIUrl":null,"url":null,"abstract":"<div><h3>Background and objective:</h3><div>Age estimation is crucial in pediatrics, developmental medicine, often conducted by radiographic techniques exposing children to ionizing radiation. Magnetic Resonance Imaging (MRI) offers a safer, radiation-free alternative. Automatic age estimation is rapidly advancing, offering an efficient approach that reduces human bias and saves manpower. This study aims to exploit the potential of head MRI in automatic age estimation in pediatric population via deep learning.</div></div><div><h3>Methods and materials:</h3><div>We propose a self-distillation and hybrid-attention network (SDHA) to estimate age from 3-T head MRI from children. We train SDHA network utilizing self-distillation and integrating Squeeze-and-Excitation (SE), Spatial Transformer (ST) attention mechanisms. Four stacked attention modules (SE, ST) were embedded to backbone network ResNet50 (teacher), generating deeper predictions; early exit branches (students) were added to generate shallower predictions. Three types of losses are employed to achieve knowledge distillation to enhance both performance and computational efficiency. SDHA is evaluated against manual and traditional CNN methods by mean absolute error (MAE) and root mean squared error (RMSE).</div></div><div><h3>Results:</h3><div>SDHA (MAE = 0.34 years) yielded a lower MAE than manual method (MAE = 0.44 years). MAE decreased by 63.4% with SDHA compared to non-distilled SENet (MAE = 0.93 years). Prediction error density curve shows higher precision by SDHA. Grad-CAM visualization revealed that SDHA adaptively focuses on age-relevant dental, facial and brain structures. SDHA reduced prediction time from 120 s (manual assessment) to 0.11 s per subject.</div></div><div><h3>Conclusion:</h3><div>The proposed SDHA demonstrates superior performance over manual and existing CNN methods for dental age estimation from head MRI. Its adaptive attention to age-relevant anatomical structures and significant efficiency gains make it valuable for applications in pediatric age estimation.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"112 ","pages":"Article 108748"},"PeriodicalIF":4.9000,"publicationDate":"2025-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Enhanced pediatric age estimation from head MRI via self-distillation hybrid-attention network\",\"authors\":\"Lei Shi , Xinran Huang , Wenchi Ke , Hrvoje Brkić , Yuchi Zhou , Ting Lu , Xian’e Tang , Lirong Qiu , Shuai Luo , Xingtao Zhang , Ziqi Cheng , Yushan Lin , Peixi Liao , Hu Chen , Yi Zhang , Yijiu Chen , Zhenhua Deng , Fei Fan\",\"doi\":\"10.1016/j.bspc.2025.108748\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><h3>Background and objective:</h3><div>Age estimation is crucial in pediatrics, developmental medicine, often conducted by radiographic techniques exposing children to ionizing radiation. Magnetic Resonance Imaging (MRI) offers a safer, radiation-free alternative. Automatic age estimation is rapidly advancing, offering an efficient approach that reduces human bias and saves manpower. This study aims to exploit the potential of head MRI in automatic age estimation in pediatric population via deep learning.</div></div><div><h3>Methods and materials:</h3><div>We propose a self-distillation and hybrid-attention network (SDHA) to estimate age from 3-T head MRI from children. We train SDHA network utilizing self-distillation and integrating Squeeze-and-Excitation (SE), Spatial Transformer (ST) attention mechanisms. Four stacked attention modules (SE, ST) were embedded to backbone network ResNet50 (teacher), generating deeper predictions; early exit branches (students) were added to generate shallower predictions. Three types of losses are employed to achieve knowledge distillation to enhance both performance and computational efficiency. SDHA is evaluated against manual and traditional CNN methods by mean absolute error (MAE) and root mean squared error (RMSE).</div></div><div><h3>Results:</h3><div>SDHA (MAE = 0.34 years) yielded a lower MAE than manual method (MAE = 0.44 years). MAE decreased by 63.4% with SDHA compared to non-distilled SENet (MAE = 0.93 years). Prediction error density curve shows higher precision by SDHA. Grad-CAM visualization revealed that SDHA adaptively focuses on age-relevant dental, facial and brain structures. SDHA reduced prediction time from 120 s (manual assessment) to 0.11 s per subject.</div></div><div><h3>Conclusion:</h3><div>The proposed SDHA demonstrates superior performance over manual and existing CNN methods for dental age estimation from head MRI. Its adaptive attention to age-relevant anatomical structures and significant efficiency gains make it valuable for applications in pediatric age estimation.</div></div>\",\"PeriodicalId\":55362,\"journal\":{\"name\":\"Biomedical Signal Processing and Control\",\"volume\":\"112 \",\"pages\":\"Article 108748\"},\"PeriodicalIF\":4.9000,\"publicationDate\":\"2025-10-09\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Biomedical Signal Processing and Control\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S1746809425012595\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"ENGINEERING, BIOMEDICAL\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Biomedical Signal Processing and Control","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1746809425012595","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, BIOMEDICAL","Score":null,"Total":0}

引用次数: 0

摘要

背景和目的：年龄估计在儿科和发育医学中是至关重要的，通常通过使儿童暴露于电离辐射的放射技术进行。磁共振成像（MRI）提供了一种更安全、无辐射的替代方法。自动年龄估计正在迅速发展，提供了一种有效的方法，减少了人类的偏见，节省了人力。本研究旨在通过深度学习挖掘头部MRI在儿童人群年龄自动估计中的潜力。方法和材料：我们提出了一种自蒸馏和混合注意网络（SDHA）来估计儿童3-T头部MRI的年龄。我们利用自蒸馏和整合挤压-激励（SE）、空间变压器（ST）注意力机制来训练SDHA网络。在骨干网络ResNet50（教师）中嵌入4个堆叠的注意力模块（SE、ST），产生更深层次的预测；早期退出分支（学生）的加入是为了产生较浅的预测。采用三种类型的损失来实现知识蒸馏，以提高性能和计算效率。通过平均绝对误差（MAE）和均方根误差（RMSE）对人工和传统CNN方法的SDHA进行评估。结果：SDHA法（MAE = 0.34年）的MAE低于手工法（MAE = 0.44年）。与未蒸馏的SENet相比，SDHA使MAE降低了63.4% （MAE = 0.93年）。SDHA预测误差密度曲线具有较高的精度。Grad-CAM可视化显示，SDHA自适应地关注与年龄相关的牙齿、面部和大脑结构。SDHA将每个受试者的预测时间从120秒（人工评估）减少到0.11秒。结论：本文提出的SDHA在头部MRI牙齿年龄估计方面优于人工和现有的CNN方法。它对年龄相关解剖结构的适应性关注和显著的效率提高使其在儿童年龄估计中的应用具有价值。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

查看原文本刊更多论文

Enhanced pediatric age estimation from head MRI via self-distillation hybrid-attention network

Background and objective:

Age estimation is crucial in pediatrics, developmental medicine, often conducted by radiographic techniques exposing children to ionizing radiation. Magnetic Resonance Imaging (MRI) offers a safer, radiation-free alternative. Automatic age estimation is rapidly advancing, offering an efficient approach that reduces human bias and saves manpower. This study aims to exploit the potential of head MRI in automatic age estimation in pediatric population via deep learning.

Methods and materials:

We propose a self-distillation and hybrid-attention network (SDHA) to estimate age from 3-T head MRI from children. We train SDHA network utilizing self-distillation and integrating Squeeze-and-Excitation (SE), Spatial Transformer (ST) attention mechanisms. Four stacked attention modules (SE, ST) were embedded to backbone network ResNet50 (teacher), generating deeper predictions; early exit branches (students) were added to generate shallower predictions. Three types of losses are employed to achieve knowledge distillation to enhance both performance and computational efficiency. SDHA is evaluated against manual and traditional CNN methods by mean absolute error (MAE) and root mean squared error (RMSE).

Results:

SDHA (MAE = 0.34 years) yielded a lower MAE than manual method (MAE = 0.44 years). MAE decreased by 63.4% with SDHA compared to non-distilled SENet (MAE = 0.93 years). Prediction error density curve shows higher precision by SDHA. Grad-CAM visualization revealed that SDHA adaptively focuses on age-relevant dental, facial and brain structures. SDHA reduced prediction time from 120 s (manual assessment) to 0.11 s per subject.

Conclusion:

The proposed SDHA demonstrates superior performance over manual and existing CNN methods for dental age estimation from head MRI. Its adaptive attention to age-relevant anatomical structures and significant efficiency gains make it valuable for applications in pediatric age estimation.

求助全文

通过发布文献求助，成功后即可免费获取论文全文。去求助

来源期刊

Biomedical Signal Processing and Control 工程技术-工程：生物医学

CiteScore

9.80

自引率

13.70%

发文量

822

审稿时长

4 months

期刊介绍： Biomedical Signal Processing and Control aims to provide a cross-disciplinary international forum for the interchange of information on research in the measurement and analysis of signals and images in clinical medicine and the biological sciences. Emphasis is placed on contributions dealing with the practical, applications-led research on the use of methods and devices in clinical diagnosis, patient monitoring and management. Biomedical Signal Processing and Control reflects the main areas in which these methods are being used and developed at the interface of both engineering and clinical science. The scope of the journal is defined to include relevant review papers, technical notes, short communications and letters. Tutorial papers and special issues will also be published.