PO.BCS01.08 · 生物信息与计算
基于手机的便携式切片数字化系统，用于支持AI的组织病理学分析
Phone-based portable slide digitization system for AI-enabled histopathology
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
在传统的组织学分析中，病理学家使用光学显微镜观察玻璃切片。如今，切片正越来越多地通过高分辨率扫描仪进行数字化。数字病理学在缓解获取专科肿瘤诊疗资源不平等方面具有巨大潜力，并可应用基于人工智能（AI）的工具进行分析，但数字化成本可能高得令人望而却步。我们开发了一套基于手机的系统，利用视频采集和图像拼接生成全切片图像，从而进行数字化、可视化并实现分析自动化，克服了手机镜头视野固有的局限性。我们的系统无需专用设备即可提供数字病理学能力。将装有显微镜镜头的手机放置在自制的3D打印支架上，支架带有光源，可对透射照明的切片进行流畅的视频采集。我们使用Apple iPhone 16 pro + Sandmarc镜头验证了系统效能。The Telepath是我们原创的应用程序，采用5倍光学变焦、锁定曝光、固定白平衡和连续自动对焦，以每秒60帧的速度采集视频，保持一致的光照和色彩。通过尺度不变特征变换（SIFT）描述符和快速近似最近邻库（FLANN）匹配器提取并对齐重叠图像。经过光照校正后，图像以TIFF格式存储。使用CTransPath提取数值型图块级AI特征，并用于训练分类模型。The Telepath被用于数字化24张由PDX黑色素瘤标本制备的组织学切片。根据组织大小，视频采集每张切片需0.5至2分钟。最终分辨率为每像素0.57微米，平均文件大小为20 MB。病理学家审阅确认，主观图像质量足以满足审阅需求。在商用扫描仪数字化图像上训练的自动肿瘤检测模型，在The Telepath数字化的图像上进行了测试，取得了召回率0.943 ± 0.096、精确率0.763 ± 0.25、准确率0.891 ± 0.047、fpp（尺寸校正后的假阳性率）0.094 ± 0.046的结果。我们提出了一套便携、经济、支持AI的数字病理学系统，可将标本从玻璃切片转化为标注的数字图像。鉴于该系统的便携性和速度，我们设想其应用场景包括：将图像直接上传至病历以进行二次或集中病理审阅、在器官获取过程中对潜在供体器官进行自动评估，以及对手术标本进行永久切片和冰冻切片分析，用于肿瘤切缘评估和淋巴结转移检测等任务。The Telepath大幅降低了实施数字病理学的经济和时间壁垒，有望缓解获取专科诊疗资源的不平等，并为更广泛地应用基于AI的工具打开大门。
查看英文原文 English abstract
In traditional histological analysis, pathologists view glass slides using light microscopy. Increasingly, slides are being digitized with high-resolution scanners. Digital pathology has great potential to mitigate disparity in access to specialized cancer care and allows application of artificial intelligence (AI)-based tools for analysis, but digitization costs can be prohibitive. We developed a phone-based system to digitize, visualize, and automate analysis using video capture and image stitching to generate whole slide images, overcoming the innate limitation in phone lens field of view. Our system provides digital pathology capability without need for specialized equipment. A cell phone with attached microscope lens is placed on a custom, 3-D printed rig with light source for smooth video capture across transilluminated slides. We demonstrate system efficacy using Apple iPhone 16 pro + Sandmarc lens. The Telepath is our original app, using 5x optical zoom, locked exposure, fixed white balance, and continuous autofocus to capture video with consistent lighting and color at 60 frames per second. Overlapping images are extracted and aligned with Scale Invariant Feature Transform descriptors and Fast Library for Approximate Nearest Neighbor matcher. After illumination correction, images are stored in TIFF format. Numerical tile-level AI-features are extracted with CTransPath and used to train classification models. The Telepath was used to digitize 24 histology slides prepared from PDX melanoma specimens. Video capture required 0.5 to 2 minutes/slide based on tissue size. Final resolution was 0.57 microns per pixel with average file size of 20 MB. Pathologist review confirms subjective image quality is ample for review. An automatic tumor detection model trained on images digitized on a commercial scanner was tested on images digitized on The Telepath, achieving recall 0.943 ± 0.096, precision 0.763 ± 0.25, accuracy 0.891 ± 0.047, fpp (size-adjusted fpr) 0.094 ± 0.046. We present a portable, affordable, AI-enabled digital pathology system that takes specimens from glass slides to annotated digital images. Given the portability and speed of the system, we envision such applications as direct upload of images to the medical record for secondary or centralized pathology review, automated assessment of prospective donor organs during procurements, and both permanent and frozen section analysis of surgical specimens for such tasks as tumor margin assessment and detection of lymph node metastases. The Telepath dramatically decreases financial and temporal barriers to the implementation of digital pathology, with potential to mitigate disparity in access to specialty care and open the door to broader application of AI-based tools.
利益披露 Disclosure
J. C. Rubinstein, None.