目的

探讨基于视觉转换器（ViT）深度学习算法磁共振弥散张量成像（DTI）的孤独症谱系障碍（ASD）分类模型（以下简称为“ASD 分类模型”），对不同年龄段ASD 患儿的临床筛查效能。

方法

选择2017年1月至2023年7月就诊于中国人民解放军总医院及成都市妇女儿童中心医院的265例ASD 患儿（0～12 岁）为研究对象，纳入ASD 组（n=265）。选择同期在这2 家医院就诊的158例典型发育（TD）儿童（0～12岁）为对照，纳入对照组（n=158）。收集2组受试儿的颅脑DTI图像，对其34与64方向进行数据融合处理，随后基于双张量无迹卡尔曼滤波（t UKF）技术，对其全脑白质纤维进行示踪。基于谱嵌入技术生成2组受试儿纤维级TractoEmbedding图像，采用基于ViT 深度学习算法，对数据扩充后的TractoEmbedding图像进行分类，建立ASD 分类模型，并分析该模型对不同年龄段（0～2岁、≥2～4岁、≥4～6岁、≥6～8岁、≥8～10岁、≥10～12岁）受试儿的ASD 筛查效能，包括准确率、特异度、敏感度及曲线下面积（AUC）。本研究通过中国人民解放军总医院医学伦理委员会审批（批准文号：S2022-646-01），与所有受试儿监护人签署临床研究知情同意书。ASD 组≥2～4岁、≥4～6岁、≥6～8岁、≥8～10岁、≥10～12岁患儿年龄分别与对照组儿童比较，差异均无统计学意义（P ＞0.05）。

结果

本研究基于ViT 深度学习算法，通过对ASD 患儿与TD 儿童纤维级TractoEmbedding图像的训练与验证后，建立5 个ASD 分类模型（模型1～5）。其中，模型3对ASD 患儿的筛查准确率最高，为0.791。采用模型3对不同年龄段受试儿ASD 筛查的准确率、特异度、敏感度及AUC值分别如下。≥4～5岁：0.966、0.864、1.000及0.961；≥5～6岁：0.891、0.842、0.926 及0.961；≥6～7 岁：0.958、0.903、1.000 及0.992；≥7～8 岁：0.893、0.850、1.000及0.956。

结论

本研究建立的基于ViT 深度学习算法DTI 的ASD 分类模型中，模型3对≥4～5岁、≥5～6岁、≥6～7 岁、≥7～8 岁受试儿ASD 筛查的敏感度、特异度、准确率、AUC 值均较高。

Objective

To explore the clinical screening efficacy of children with autism spectrum disorder（ASD）in different ages by ASD classification model based on the visual transformer（ViT）deep learning algorithm diffusion tensor imaging （DTI） （hereinafter referred to as the"ASD classification model"）.

Methods

A total of 265 ASD children aged 0-12 years who visited the Chinese PLA General Hospital and the Chengdu Women's and Children's Central Hospital from January 2017 to July 2023 were selected into ASD group，And other 158 typical development（TD）children aged 0-12 years old who visited the above two hospitals during the same period were included in the control group.Brain DTIimages of all children were collected，and the DTIimage data of 34 and 64 directions of children from the two hospitals were harmonized，and then whole brain white matter fibers of the children were traced based on two-tensor unscented Kalman filter （t UKF）technology.Based on spectral embedding technology，fiber level TractoEmbedding images of the children were generated.Then augmented TractoEmbedding images were classified using ViT deep learning algorithm，and ASD classification model was established，and the screening performance of the model for ASD with different ages（0-2 years old，≥2-4 years old，≥4-6 years old，≥6-8 years old，≥8-10 years old，≥10-12 years old）was analyzed，including accuracy，specificity，sensitivity，and area under the curve（AUC）.This study was approved by the Medical Ethics Committee of Chinese PLA General Hospital（Approval No.S2022-646-01）.All guardians of the enrolled children signed the informed consent form for clinical study.There was no statistically s ignificant difference in the age of children aged≥2-4 years，≥4-6 years，≥6-8 years，≥8-10 years，and≥10-12 years between ASD group and control group （P＞0.05）.

Results

Based on the ViT deep learning algorithm，this study established five ASD classification models （Models 1-5）by training and validating fiber level DTI TractoEmbedding images of ASD and TD children.Among them，Model 3 had the highest screening accuracy for ASD at 0.791.The results of screening performance in children of different ages by Model 3 showed that the accuracy，specificity，sensitivity，and AUC values of the model for ASD screening in children aged≥4-5 years，≥5-6 years，≥6-7 years，and≥7-8 years were as follows.≥4-5 years old：0.966，0.864，1.000，0.961；≥5-6 years old：0.891，0.842，0.926，0.961；≥6-7 years old：0.958，0.903，1.000，0.992；≥7-8 years old：0.893，0.850，1.000，0.956.

Conclusions

The ASD classification model 3 based on ViT deep learning algorithm DTI established in this study has high sensitivity，specificity，accuracy，and AUC values for ASD screening in children aged≥4-5 years，≥5-6 years，≥6-7 years，and≥7-8 years.