PO.TB04.05 · 肿瘤生物学
从数小时到数秒：一种利用AI算法加速体内啮齿动物成像数据分析的新工具
From hours to seconds: a new tool for accelerating in vivo rodent imaging data analysis using AI algorithms
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：癌症研究中无创成像技术的进步增加了对更大动物队列的需求，以达到统计效力。然而，体内图像分析仍然具有挑战性，通常需要熟练的用户在二维和三维图像中手动处理器官或肿瘤——这是一项耗时的任务，视研究规模可能需要数小时到数周。为不同成像模态掌握多种软件平台的需求进一步延长了分析时间，往往超过图像采集时间。
方法：为应对这些挑战，我们开发了一款基于Python的多模态软件应用程序，旨在通过AI辅助分割和跨多个时间点的批量分析来加速数据处理。在此，我们报告该软件以及多个集成AI分割模型在超声和光学成像方面的性能。
结果：与手动工作流程相比，三维超声的分析通量提高了9倍，二维光学成像的分析通量提高了最多60倍。该软件相比人工真值分割和离体验证标准表现出高度一致性。
BLI成像：基于深度学习的掩模处理与标准定量方法几乎完全一致（R^2 = 0.995 对比圆形ROI；R^2 = 0.996 对比边界框），同时将每次研究的分析时间从15-20秒缩短至约2秒。
超声成像：AI测量的脾脏大小与死后脾脏重量（R^2 = 0.93）和MRI体积（R^2 = 0.90）高度相关。AI分割相对于人工真值分割达到平均Dice分数0.89，预测体积的相关性为R^2 = 0.95。对于皮下肿瘤，尽管其回声纹理更为异质、更具挑战性，AI与人工分割相比的一致性较低，但仍然较强（Dice = 0.82；R^2 = 0.78）。
结论：AI驱动的自动化在不影响准确性的前提下显著加速了多模态图像分析。这些进展突显了集成自动化在简化临床前成像工作流程和提高研究效率方面的潜力。
查看英文原文 English abstract
Background: Advances in noninvasive imaging for cancer research have increased the demand for larger animal cohorts to achieve statistical power. However, in vivo image analysis remains challenging, often requiring skilled users to manually process organs or tumors in 2D and 3D images-a time-intensive task that can take hours to weeks depending on study size. The need to master multiple software platforms for different imaging modalities further extends analysis time, often surpassing image acquisition time.
Methods: To address these challenges, we developed a Python-based multimodal software application designed to accelerate data processing through AI-assisted segmentation and batch analysis across multiple timepoints. Here, we report the performance of the software and multiple integrated AI segmentation models for ultrasound and optical imaging.
Results: Analysis throughput improved by 9x for 3D ultrasound and up to 60x for 2D optical imaging compared to manual workflows. The software demonstrated strong agreement compared to human ground truth segmentations and ex vivo validation standards.
BLI Imaging: Deep learning-based masking showed near-perfect agreement with standard quantification methods (R^2 = 0.995 vs circular ROI; R^2 = 0.996 vs bounding boxes) while reducing analysis time from 15-20 seconds to ~2 seconds per study.
Ultrasound Imaging: AI-measured spleen size correlated strongly with postmortem spleen weights (R^2 = 0.93) and MRI volumes (R^2 = 0.90). AI segmentations achieved an average Dice score of 0.89 against ground truth human segmentations with predicted volumes correlating at R^2 = 0.95. For subcutaneous tumors, agreement was lower, but still strong when comparing AI versus human segmentations (Dice = 0.82; R^2 = 0.78) despite a more challenging heterogeneous echotexture profile.
Conclusions: AI-driven automation significantly accelerates multimodal image analysis without compromising accuracy. These advances highlight the potential of integrated automation to streamline preclinical imaging workflows and enhance research efficiency.
利益披露 Disclosure
H. Sweezo,
Revvity Employment.
J. Rojas,
Revvity Employment.
T. Kierski,
Revvity Employment.
A. Aji,
Revvity Employment.
J. Pesner,
Revvity Employment.
J. Betthauser,
Revvity Employment.
Z. Houston,
Revvity Employment.
K. Kloepping,
Revvity Employment.
J. Tseng,
Revvity Employment.
C. McMannus,
Revvity Employment.
B. John,
Revvity Employment.
J. Peterson,
Revvity Employment.
J. B. Schueler,
Charles River Labs Employment.
R. Gessner,
Revvity Employment.
T. Czernuszewicz,
Revvity Employment.