Abstract Background. Fluorescent imaging technologies now allow for multiplex immunofluorescence (mIF) for up to 100 targets on a single slide. However, the ability to quantitatively analyze the resulting data, especially on whole-slide images (WSI), is limited by scalability and reproducibility. Currently available platforms for segmenting cancer cells and nuclei involve segmentation algorithms that are hand-tuned on individual fields of view, making these methods subjective and difficult to replicate. To this end, we sought to develop an end-to-end workflow for WSI mIF data in cancer, from raw images to cell-level features, using state-of-the-art deep learning models for tissue, cell, and nuclei segmentation. Methods. mIF was performed using the Akoya 6-plex Lung IO panel (CD8, FoxP3, CD68, PD-1, PD-L1, and pan-cytokeratin) with a DAPI counterstain on clinical NSCLC specimens obtained from commercial sources (N=41). Slides were scanned using the Akoya PhenoImager HT. To remove bleed-through between different fluorophores, mIF WSIs were linearly unmixed based on a reference single-stain matrix, generated by selecting pure stain regions in mIF images. Convolutional neural network models previously trained on hematoxylin and eosin (H&E) images for 1) artifact detection, 2) tissue region identification, and 3) nucleus detection and segmentation were then deployed on mIF images that were converted to RGB images using a synthetic H&E transformation. Cell segmentation models were developed and implemented for each marker assessed. For each demixed channel representing a single antibody signal, expert annotations were used to train deep learning models for cell detection and segmentation. Finally, segmentation results from separate channels were aggregated based on spatial colocalization to identify cells and cell phenotypes. Results. We developed a pipeline for mIF image analysis at scale. After unmixing, conversion of an mIF image to a synthetic H&E image allowed for 1) removal of artifact regions, 2) detection of cancer and stroma using tissue segmentation, and 3) nuclei segmentation. Nuclear segmentation using this approach improved upon a widely used commercial segmentation platform when compared to manual expert annotations (Dice scores = 0.85 and 0.52, respectively). Finally, using state-of-the-art instance segmentation models allows a novel approach to cell phenotyping compared to currently available platforms, allowing positive identification of cells lacking a nuclear signal in the sectioned plane (e.g. large cytokeratin-positive cells). Conclusions. We developed a robust workflow to accurately segment cells and nuclei from mIF WSIs at scale. Improving the scalability and reproducibility of complex mIF image analysis may help improve the adoption of this approach for biomarker development and incorporation into clinical trial workflows in oncology. Citation Format: Waleed Tahir, Emma Krause, Jin Li, Howard Mak, Mohammad Mirzadeh, Kevin Rose, Judy Shen, Vignesh Valaboju, Guillaume Chhor, Joseph Lee, Jun Zhang, Jacqueline Brosnan-Cashman, Michael G. Drage, Justin Lee, Carlee Hemphill, Saumya Pant, Robert Egger. Development of a high-throughput image processing pipeline for multiplex immunofluorescence whole slide images at scale [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 6616.
Read full abstract
Sep 27, 2024
Masahide Anada + 8
Open Access