Lou Jerry J, Chang Peter, Nava Kiana D, Chantaduly Chanon, Wang Hsin-Pei, Yong William H, Patel Viharkumar, Chaudhari Ajinkya J, Vasquez La Rissa, Monuki Edwin, Head Elizabeth, Vinters Harry V, Magaki Shino, Harvey Danielle J, Chuah Chen-Nee, DeCarli Charles S, Williams Christopher K, Keiser Michael, Dugger Brittany N
Department of Pathology and Laboratory Medicine, School of Medicine, University of California Irvine, Irvine, USA.
Department of Radiological Sciences, Center for Artificial Intelligence in Diagnostic Medicine, School of Medicine, University of California Irvine, Orange, USA.
Free Neuropathol. 2025 Jun 2;6:12. doi: 10.17879/freeneuropathology-2025-6387. eCollection 2025.
Objective quantification of brain arteriolosclerosis remains an area of ongoing refinement in neuropathology, with current methods primarily utilizing semi-quantitative scales completed through manual histological examination. These approaches offer modest inter-rater reliability and do not provide precise quantitative metrics. To address this gap, we present a prototype end-to-end machine learning (ML)-based algorithm, Arteriolosclerosis Segmentation (ArtSeg), followed by Vascular Morphometry (VasMorph) - to assist persons in the morphometric analysis of arteriolosclerotic vessels on whole slide images (WSIs). We digitized hematoxylin and eosin-stained glass slides (13 participants, total 42 WSIs) of human brain frontal or occipital lobe cortical and/or periventricular white matter collected from three brain banks (University of California, Davis, Irvine, and Los Angeles Alzheimer's Disease Research Centers). ArtSeg comprises three ML models for blood vessel detection, arteriolosclerosis classification, and segmentation of arteriolosclerotic vessel walls and lumens. For blood vessel detection, ArtSeg achieved area under the receiver operating characteristic curve (AUC-ROC) values of 0.79 (internal hold-out testing) and 0.77 (external testing), Dice scores of 0.56 (internal hold-out) and 0.74 (external), and Hausdorff distances of 2.53 (internal hold-out) and 2.15 (external). Arteriolosclerosis classification demonstrated accuracies of 0.94 (mean, 3-fold cross-validation), 0.86 (internal hold-out), and 0.77 (external), alongside AUC-ROC values of 0.69 (mean, 3-fold cross-validation), 0.87 (internal hold-out), and 0.83 (external). For arteriolosclerotic vessel segmentation, ArtSeg yielded Dice scores of 0.68 (mean, 3-fold cross-validation), 0.73 (internal hold-out), and 0.71 (external); Hausdorff distances of 7.63 (mean, 3-fold cross-validation), 6.93 (internal hold-out), and 7.80 (external); and AUC-ROC values of 0.90 (mean, 3-fold cross-validation), 0.92 (internal hold-out), and 0.87 (external). VasMorph successfully derived sclerotic indices, vessel wall thicknesses, and vessel wall to lumen area ratios from ArtSeg-segmented vessels, producing results comparable to expert assessment. This integrated approach shows promise as an assistive tool to enhance current neuropathological evaluation of brain arteriolosclerosis, offering potential for improved inter-rater reliability and quantification.
脑小动脉硬化的客观量化仍是神经病理学中一个不断完善的领域，目前的方法主要是利用通过手动组织学检查完成的半定量量表。这些方法的评分者间信度一般，且无法提供精确的定量指标。为弥补这一差距，我们提出了一种基于机器学习（ML）的端到端原型算法，即小动脉硬化分割（ArtSeg）算法，随后是血管形态测量（VasMorph）算法，以辅助人员对全玻片图像（WSIs）上的小动脉硬化血管进行形态测量分析。我们将从三个脑库（加利福尼亚大学戴维斯分校、尔湾分校和洛杉矶分校的阿尔茨海默病研究中心）收集的人脑额叶或枕叶皮质和/或脑室周围白质的苏木精和伊红染色玻片（13名参与者，共42张WSIs）进行了数字化处理。ArtSeg算法包含三个用于血管检测、小动脉硬化分类以及小动脉硬化血管壁和管腔分割的ML模型。对于血管检测，ArtSeg算法在内部保留测试中的受试者工作特征曲线下面积（AUC-ROC）值为0.79，在外部测试中为0.77；Dice分数在内部保留测试中为0.56，在外部测试中为0.74；豪斯多夫距离在内部保留测试中为2.53，在外部测试中为2.15。小动脉硬化分类的准确率在平均3折交叉验证中为0.94，在内部保留测试中为0.86，在外部测试中为0.77，同时AUC-ROC值在平均3折交叉验证中为0.69，在内部保留测试中为0.87，在外部测试中为0.83。对于小动脉硬化血管分割，ArtSeg算法的Dice分数在平均3折交叉验证中为0.68，在内部保留测试中为0.73，在外部测试中为0.71；豪斯多夫距离在平均3折交叉验证中为7.63，在内部保留测试中为6.93，在外部测试中为7.80；AUC-ROC值在平均3折交叉验证中为0.90，在内部保留测试中为0.92，在外部测试中为0.87。VasMorph算法成功地从ArtSeg算法分割的血管中得出硬化指数、血管壁厚度以及血管壁与管腔面积比，其结果与专家评估相当。这种综合方法有望成为一种辅助工具，以加强目前对脑小动脉硬化的神经病理学评估，提高评分者间信度和量化水平。
