PO.BCS01.07 · 生物信息与计算

准确的焦平面对于AI评估用于膀胱癌筛查和监测的非单层尿液细胞学标本至关重要

Accurate focal plane is crucial for AI assessment of non-monolayer urine cytology specimens for bladder cancer screening and surveillance

编号 1446 展板 9 时间 4/20 09:00–12:00 区域 Section 4 主讲 Brody McNutt, BA;MS

分会场 Digital Pathology 2

查看 PDF 下载 PDF 🔒 查看 / 下载完整 PDF 需登录并开通下载套餐 · 查看套餐 / 开通 AACR 官方页面

作者与单位 Authors & Affiliations

Brody McNutt¹, Sam Harvey², Minh-Khang Le¹, I-Chuang Liao¹, Keluo Yao¹, Xiaoying Liu³, Camille Ng¹, Ahmad Kohsar², Daniel Shou², Christopher VandenBussche², Louis J. Vaickus³, Joshua Jay Levy¹

¹Cedars-Sinai Medical Center, Los Angeles, CA,²Johns Hopkins Medical Institutions, Baltimore, MD,³Dartmouth-Hitchcock Medical Center, Lebanon, NH

摘要 Abstract

中文摘要

背景：膀胱癌需要频繁监测，尿液细胞学被广泛用于指导膀胱镜评估。数字图像分析旨在提供与Paris系统一致的定量细胞级指标，但许多实验室使用非单层制片(如SurePath)，这些制片将细胞置于三维排列中，使自动化评估复杂化。我们评估了焦平面选择如何影响核质比(NC比)估计，并比较了几种用于识别最佳焦平面的算法策略。方法：我们分析了在Roche Ventana DP 200系统(Johns Hopkins)上扫描的300张SurePath全切片图像，均匀涵盖阴性、非典型、可疑和高级别癌病例。已发表的检测模型识别出细胞簇，经六名病理医生重新标注，产生343个细胞簇和2,435个尿路上皮细胞。标注者在每个细胞的最佳对焦平面勾勒细胞核和细胞质区域。我们评估了经典对焦指标(改进拉普拉斯之和[SML]、高频能量[HFE]、Tenengrad、Brenner梯度、Laplacian以及基于熵的OpenCV方法)和无监督视觉transformer方法(特征方差[ViT-V]、注意力熵[ViT-A]以及直接预测焦平面的监督式Z-stack transformer模型[ViT-T、ViT-CLS])。算法通过相对于病理医生真实标准的1平面内准确率进行评估。NC比由核/质区域推导得出。U-Net分割为留出测试集生成NC比，并采用以下方式评估与真实标准的Spearman相关性：(1)病理医生选择的平面；(2)离焦平面图像；(3)算法选择的平面。结果：1平面内准确率为0.416(ViT-A)、0.594(Grad)、0.740(HFE)、0.779(OpenCV)、0.789(ViT-V)、0.853(Laplacian)、0.857(Tenengrad)、0.862(SML)，Z-stack transformer ViT-T和ViT-CLS最高(0.874、0.872)。在病理医生选择的平面上NC比估计的相关性达到0.774。相关性随图像离焦而下降(±1平面时约0.74；±2时约0.69；±3时约0.65；±4时约0.59；±5时约0.50)。使用算法选择的平面时，相关性为0.639(ViT-A)、0.687(Grad)、0.720(ViT-V)、0.728(HFE)、0.733(OpenCV)、0.743(Tenengrad)、0.744(Laplacian)、0.746(SML)以及0.745/0.738(ViT-T/ViT-CLS)。结论：准确的焦平面选择对于非单层尿液制片中可靠的基于AI的细胞学评估至关重要。NC比准确性和下游分析有效性在离焦时迅速退化，而算法选择的平面恢复了这一损失的大部分。未来的工作将评估其对细胞簇级和患者级任务的影响，并评估将z-stack中最清晰区域拼接成单一最佳对焦图像的扩展景深融合方法。

查看英文原文 English abstract

Background: Bladder cancer requires frequent surveillance, and urine cytology is widely used to guide cystoscopic evaluation. Digital image analysis aims to provide quantitative cell-level metrics aligned with The Paris System, but many laboratories use non-monolayer preparations (e.g., SurePath) that place cells in three-dimensional arrangements, complicating automated evaluation. We assessed how focal-plane selection affects nuclear-to-cytoplasmic (NC) ratio estimation and compared several algorithmic strategies for identifying the optimal focal plane. Methods: We analyzed 300 SurePath whole-slide images scanned on a Roche Ventana DP 200 system (Johns Hopkins), evenly spanning negative, atypical, suspicious, and high-grade carcinoma cases. A published detection model identified clusters, which were reannotated by six pathologists, yielding 343 clusters and 2,435 urothelial cells. Annotators outlined nuclei and cytoplasm areas at each cell's best-focus plane. We evaluated classical focus metrics (Sum of Modified Laplacian [SML], High-Frequency Energy [HFE], Tenengrad, Brenner Gradient, Laplacian, and entropy-based OpenCV methods) and unsupervised vision-transformer approaches (feature-variance [ViT-V], attention-entropy [ViT-A], and supervised Z-stack transformer models [ViT-T, ViT-CLS] that directly predict the focal plane). Algorithms were assessed by within-1-plane accuracy relative to pathologist ground truth. NC ratios were derived from nuclear/cytoplasmic areas. U-Net segmentation generated NC ratios for a held-out test set, and Spearman correlations with ground truth were evaluated using: (1) pathologist-selected planes; (2) off-plane images; and (3) algorithm-selected planes. Results: Within-1-plane accuracy was 0.416 (ViT-A), 0.594 (Grad), 0.740 (HFE), 0.779 (OpenCV), 0.789 (ViT-V), 0.853 (Laplacian), 0.857 (Tenengrad), 0.862 (SML), and highest for Z-stack transformers ViT-T and ViT-CLS (0.874, 0.872). NC-ratio estimation at the pathologist-selected plane reached a correlation of 0.774. Correlations decreased as images moved off-plane (~0.74 at ±1 plane; ~0.69 at ±2; ~0.65 at ±3; ~0.59 at ±4; ~0.50 at ±5). Using algorithm-selected planes, correlations were 0.639 (ViT-A), 0.687 (Grad), 0.720 (ViT-V), 0.728 (HFE), 0.733 (OpenCV), 0.743 (Tenengrad), 0.744 (Laplacian), 0.746 (SML), and 0.745/0.738 (ViT-T/ViT-CLS). Conclusion: Accurate focal-plane selection is essential for reliable AI-based cytologic assessment in non-monolayer urine preparations. NC-ratio accuracy and downstream analytic validity degraded quickly off-plane, while algorithm-selected planes recovered much of this loss. Future work will evaluate impacts on cluster- and patient-level tasks and assess extended-focus fusion methods that stitch the sharpest regions across the z-stack into a single optimally focused image.

利益披露 Disclosure

B. McNutt, None.. S. Harvey, None.. M. Le, None.. I. Liao, None.. K. Yao, None.. X. Liu, None.. C. Ng, None.. A. Kohsar, None.. D. Shou, None.. C. VandenBussche, None.. L. J. Vaickus, None.. J. J. Levy, None.

← 返回 AACR 2026 检索