A Study on an Object Detection Model for Detecting Acute Small-Area Infarct Lesions on NCCT

doi:10.3969/j.issn.1673-5765.2026.01.009

Abstract

Abstract: Objective To develop an object detection model for the automatic detection of acute small-area infarct lesions on non-contrast CT (NCCT) based on an improved YOLOv5 deep learning model.
Methods Patients with acute ischemic stroke who were admitted to the Affiliated Hospital of Yangzhou University from January 2018 to December 2023 were enrolled in this retrospective study. They were randomly divided into a training/validation set, and a test set at a ratio of 10∶1. The training/validation set was used for model parameter fitting, applied to compare different loss functions for optimal model selection, and the detection performance of the model was evaluated using precision, recall, and mean average precision (mAP). The test set was utilized for the independent assessment of diagnostic efficacy. MRI DWI images were registered with NCCT images, and the lesions were labeled on the NCCT images. With the DWI-Alberta stroke program early CT score (ASPECTS) as the gold standard, the sensitivity, specificity, and accuracy of the model and physicians in detecting infarct lesions in each ASPECTS brain region were calculated respectively in the test set. The McNemar test was adopted to compare the differences in diagnostic efficacy between the model and physicians. The intra-class correlation coefficient (ICC) and weighted Kappa test were used to evaluate the consistency of CT-ASPECTS derived from the model/physicians and the gold standard DWI-ASPECTS. Meanwhile, the Bootstrap resampling method was used to test the difference between the two consistency coefficients.
Results A total of 275 patients with acute ischemic stroke were enrolled, including 250 patients in the training/validation set, and 25 patients in the test set. The improved YOLOv5 deep learning model achieved the optimal performance during the training/validation phase, with a precision of 0.824, a recall of 0.810, and an mAP@0.5 of 0.785. In the test set, comparison results of lesion detection efficacy showed that the overall accuracy (96.00% vs. 91.11%) and specificity (98.74% vs. 94.70%) of the model group were significantly superior to the physician group (both P<0.001). The model group showed a trend toward higher sensitivity compared to the physician group (75.93% vs. 64.81%), but the difference was not statistically significant (P=0.288). In the subgroup analysis of each brain region, the accuracy of the model group in the M6 brain region was higher than that of the physician group (98.00% vs. 84.00%, P=0.039). The consistency between the CT-ASPECTS of the model group and the gold standard DWI-ASPECTS (ICC 0.669, P<0.001; weighted κ=0.447, P<0.001) tended to be higher than that between CT-ASPECTS of the physician group and the gold standard DWI-ASPECTS (ICC 0.452, P=0.010; weighted κ=0.247, P=0.054). Bootstrap analysis showed that the difference in weighted κ between the model group and the physician group was statistically significant (P=0.044).
Conclusions The object detection model constructed in this study can realize automatic detection of acute small-area infarct lesions on NCCT images of patients with acute ischemic stroke. It is conducive to reducing missed diagnoses, improving the efficiency of imaging diagnosis, and providing a reliable auxiliary tool for clinical practice.

Key words: Acute ischemic stroke; Infarct core; Deep learning; Artificial intelligence; Alberta stroke program early computed tomography score

摘要： 目的基于改进的YOLOv5深度学习模型，构建在CT平扫（non-contrast CT，NCCT）上自动检测急性小面积梗死灶的目标检测模型。
方法回顾性纳入2018年1月—2023年12月于扬州大学附属医院就诊的急性缺血性卒中患者，按10∶1的比例随机分为训练/验证集与测试集。训练/验证集用于模型参数拟合，比较不同损失函数以优选模型，并采用精确率、召回率及平均精度（mean average precision，mAP）评估模型的检测效能；测试集用于独立评估模型的诊断效能。将MRI DWI图像与NCCT图像进行配准，并在NCCT图像上标记病灶。以DWI-Alberta卒中项目早期CT评分（Alberta stroke program early CT score，ASPECTS）为金标准，在测试集中分别计算模型及医师对ASPECTS各脑区梗死灶检出的敏感度、特异度及准确度，通过McNemar检验比较模型与医师的诊断效能差异。采用组内相关系数（intra-class correlation coefficient，ICC）与加权Kappa检验评估模型及医师CT-ASPECTS与金标准DWI-ASPECTS之间的一致性，并通过Bootstrap自助抽样法对两者的一致性系数进行差异性检验。
结果共纳入急性缺血性卒中患者275例，其中训练/验证集250例，测试集25例。改进的YOLOv5深度学习模型在训练/验证阶段的性能最佳：精确率为0.824，召回率为0.810，mAP@0.5为0.785。测试集中，模型病灶检测效能的比较结果显示，模型组的总体准确度（96.00% vs. 91.11%）及特异度（98.74% vs. 94.70%）优于医师组（均P<0.001）；模型组敏感度有高于医师组的趋势（75.93% vs. 64.81%），但差异无统计学意义（P=0.288）。在各脑区的亚组分析中，模型组在M6脑区的准确度高于医师组（98.00% vs. 84.00%，P=0.039）。模型组CT-ASPECTS与金标准DWI-ASPECTS的一致性（ICC 0.669，P<0.001；加权κ=0.447，P<0.001）较医师组CT-ASPECTS与金标准DWI-ASPECTS的一致性（ICC 0.452，P=0.010；加权κ=0.247，P=0.054）有提高趋势；Bootstrap分析显示，模型组与医师组加权κ的差异具有统计学意义（P=0.044）。
结论本研究构建的目标检测模型可在急性缺血性卒中患者的NCCT图像上实现急性小面积梗死灶的自动检测，有助于减少漏诊、提高影像诊断效率，为临床提供可靠的辅助工具。

关键词: 急性缺血性卒中; 梗死核心; 深度学习; 人工智能; Alberta卒中项目早期计算机断层扫描评分

CLC Number:

BAN Qiqi, WANG Wei, GUO Yuan, LI Yang, SU Xingyue, QU Hang. A Study on an Object Detection Model for Detecting Acute Small-Area Infarct Lesions on NCCT[J]. Chinese Journal of Stroke, 2026, 21(1): 78-87.

班淇琦, 王苇, 郭远, 厉阳, 苏星月, 瞿航. 构建目标检测模型在NCCT上检测急性小面积梗死灶的研究[J]. 中国卒中杂志, 2026, 21(1): 78-87.

References

[1] GBD 2019 Stroke Collaborators. Global，regional，and national burden of stroke and its risk factors，1990-2019：a systematic analysis for the global burden of disease study 2019[J]. Lancet Neurol，2021，20（10）：795-820.
[2] XING L Y，JING L，TIAN Y M，et al. High prevalence of stroke and uncontrolled associated risk factors are major public health challenges in rural northeast China：a population-based study[J]. Int J Stroke，2020，15（4）：399-411.
[3] QIU W，KUANG H L，TELEG E，et al. Machine learning for detecting early infarction in acute stroke with non-contrast-enhanced CT[J]. Radiology，2020，294（3）：638-644.
[4] 国家卫生健康委员会脑卒中防治工程委员会神经影像专业委员会，中华医学会放射学分会神经学组. 脑血管病影像规范化应用中国指南[J]. 中华放射学杂志，2019，53（11）：916-940.
National Health Commission Stroke Prevention Expert Commission Neuroimaging Committee，Chinese Medical Association Radiology Society Neuroimaging Group. Chinese guidelines for standard utilization of imaging for cerebrovascular diseases[J]. Chin J Radiol，2019，53（11）：916-940.
[5] BARBER P A，DEMCHUK A M，ZHANG J，et al. Validity and reliability of a quantitative computed tomography score in predicting outcome of hyperacute stroke before thrombolytic therapy[J]. Lancet，2000，355（9216）：1670-1674.
[6] POWERS W J，RABINSTEIN A A，ACKERSON T，et al. Guidelines for the early management of patients with acute ischemic stroke：2019 update to the 2018 guidelines for the early management of acute ischemic stroke：a guideline for healthcare professionals from the American Heart Association/American Stroke Association[J/OL]. Stroke，2019，50（12）：e344-e418[2025-08-13]. https://doi.org/10.1161/STR.0000000000000211.
[7] FARZIN B，FAHED R，GUILBERT F，et al. Early CT changes in patients admitted for thrombectomy：intrarater and interrater agreement[J]. Neurology，2016，87（3）：249-256.
[8] WINTERMARK M，REICHHART M，THIRAN J P，et al. Prognostic accuracy of cerebral blood flow measurement by perfusion computed tomography，at the time of emergency room admission，in acute stroke patients[J]. Ann Neurol，2002，51（4）：417-432.
[9] WINTERMARK M，FISCHBEIN N J，SMITH W S，et al. Accuracy of dynamic perfusion CT with deconvolution in detecting acute hemispheric stroke[J]. AJNR Am J Neuroradiol，2005，26（1）：104-112.
[10] HOELTER P，MUEHLEN I，GOELITZ P，et al. Automated ASPECT scoring in acute ischemic stroke：comparison of three software tools[J]. Neuroradiology，2020，62（10）：1231-1238.
[11] CAO Z H，XU J N，SONG B，et al. Deep learning derived automated ASPECTS on non-contrast CT scans of acute ischemic stroke patients[J]. Hum Brain Mapp，2022，43（10）：3023-3036.
[12] WEI J Y，SHANG K，WEI X E，et al. Deep learning-based automatic ASPECTS calculation can improve diagnosis efficiency in patients with acute ischemic stroke：a multicenter study[J]. Eur Radiol，2025，35（2）：627-639.
[13] PARSONS M W. Automated measurement of computed tomography acute ischemic core in stroke：does the emperor have no clothes？[J]. Stroke，2021，52（2）：642-644.
[14] FAGHANI S，CODIPILLY D C，DAVID VOGELSANG，et al. Development of a deep learning model for the histologic diagnosis of dysplasia in Barrett’s esophagus[J]. Gastrointest Endosc，2022，96（6）：918-925.
[15] ZHANG L F，CUI H Y，HU A M，et al. An improved detection algorithm for ischemic stroke NCCT based on YOLOv5[J/OL]. Diagnostics（Basel），2022，12（11）：2591[2025-08-13]. https://doi.org/10.3390/diagnostics12112591.
[16] ADAMS H P，Jr，BENDIXEN B H，KAPPELLE L J，et al. Classification of subtype of acute ischemic stroke. Definitions for use in a multicenter clinical trial. TOAST. Trial of Org 10172 in acute stroke treatment[J]. Stroke，1993，24（1）：35-41.
[17] LI Y X，TANG H，WANG W，et al. Dual attention network for unsupervised medical image registration based on VoxelMorph[J/OL]. Sci Rep，2022，12（1）：16250[2025-08-13]. https://doi.org/10.1038/s41598-022-20589-7.
[18] WANG C Y，BOCHKOVSKIY A，LIAO H Y M. Scaled-YOLOv4：scaling cross stage partial network[EB/OL].（2021-11-02）[2025-08-13]. https://ieeexplore.ieee.org/document/9577489/citations?tabFilter=papers#citations.
[19] REZATOFIGHI H，TSOI N，GWAK J Y，et al. Generalized intersection overunion：a metric and a loss for bounding box regression[EB/OL].（2020-01-09）[2025-08-13]. https://ieeexplore.ieee.org/document/8953982.
[20] ZHENG Z H，WANG P，LIU W，et al. Distance-loU loss：faster and bettter learning for bounding box regression[EB/OL].（2019-11-19）[2025-08-13]. https://arxiv.org/abs/1911.08287.
[21] CANTY A，RIPLEY B D. Boot：bootstrap R（S-Plus）functions[EB/OL].（2025-08-29）[2026-01-10]. https://cran.r-project.org/web/packages/boot/boot.pdf.
[22] DAVISON A C，HINKLEY D V. Bootstrap methods and their application[M]. Cambridge：Cambridge University Press，1997.
[23] GOYAL M，DEMCHUK A M，MENON B K，et al. Randomized assessment of rapid endovascular treatment of ischemic stroke[J]. N Engl J Med，2015，372（11）：1019-1030.
[24] MITOMI M，KIMURA K，AOKI J，et al. Comparison of CT and DWI findings in ischemic stroke patients within 3 hours of onset[J]. J Stroke Cerebrovasc Dis，2014，23（1）：37-42.
[25] VERT C，PARRA-FARIÑAS C，ROVIRA À. MR imaging in hyperacute ischemic stroke[J/OL]. Eur J Radiol，2017，96：125-132[2025-08-13]. https://doi.org/10.1016/j.ejrad.2017.06.013.
[26] GOEBEL J，STENZEL E，GUBERINA N，et al. Automated ASPECT rating：comparison between the Frontier ASPECT score software and the Brainomix software[J]. Neuroradiology，2018，60（12）：1267-1272.
[27] HERWEH C，RINGLEB P A，RAUCH G，et al. Performance of e-ASPECTS software in comparison to that of stroke physicians on assessing CT scans of acute ischemic stroke patients[J]. Int J Stroke，2016，11（4）：438-445.
[28] WOLFF L，BERKHEMER O A，VAN ES A C G M，et al. Validation of automated Alberta stroke program early CT score（ASPECTS）software for detection of early ischemic changes on non-contrast brain CT scans[J]. Neuroradiology，2021，63（4）：491-498.
[29] KUANG H L，QIU W，NAJM M，et al. Validation of an automated ASPECTS method on non-contrast computed tomography scans of acute ischemic stroke patients[J]. Int J Stroke，2020，15（5）：528-534.
[30] KUANG H L，NAJM M，CHAKRABORTY D，et al. Automated ASPECTS on noncontrast CT scans in patients with acute ischemic stroke using machine learning[J]. AJNR Am J Neuroradiol，2019，40（1）：33-38.
[31] MOKIN M，PRIMIANI C T，SIDDIQUI A H，et al. ASPECTS（Alberta stroke program early CT score）measurement using Hounsfield unit values when selecting patients for stroke thrombectomy[J]. Stroke，2017，48（6）：1574-1579.
[32] YAN Y T，CONZE P H，LAMARD M，et al. Towards improved breast mass detection using dual-view mammogram matching[J/OL]. Med Image Anal，2021，71：102083[2025-08-13]. https://doi.org/10.1016/j.media.2021.102083.
[33] LESLIE-MAZWI T M，HIRSCH J A，FALCONE G J，et al. Endovascular stroke treatment outcomes after patient selection based on magnetic resonance imaging and clinical criteria[J]. JAMA Neurol，2016，73（1）：43-49.
[34] MURRAY N M，UNBERATH M，HAGER G D，et al. Artificial intelligence to diagnose ischemic stroke and identify large vessel occlusions：a systematic review[J]. J Neurointerv Surg，2020，12（2）：156-164.
[35] BULUT H T，YILDIRIM A，EKMEKCI B，et al. False-negative diffusion-weighted imaging in acute stroke and its frequency in anterior and posterior circulation ischemia[J]. J Comput Assist Tomogr，2014，38（5）：627-633.