Abstract 5087: Single cell AI-prediction of protein biomarker on novel whole slide unstained blood smear images

Introduction: Identification of immune cell types within bodily fluids, including blood samples, plays a critical role in early detection and disease monitoring within public health initiatives. Although protein biomarkers present in blood samples are useful for clinical assessment, the process of sample preparation for immunostaining is time-consuming and costly. In this study, we present an efficient and rapid preparation procedure for affixing cells on a glass slide for downstream histological staining. Furthermore, using CD3+ T cells as a paradigm, our study concurrently investigates if deep learning models can predict these cells from unlabeled brightfield images of these cell smears. The prediction models could potentially aid in the initial screening process of blood samples and identify samples that require further testing, therefore increasing the throughput of obtaining cell identification results. Material and Methods: Buffy coat samples were separately processed and spread as monolayers across positively charged glass slides, where they were subjected to fluorescent IHC (fIHC) staining (CD3 and DAPI) using the automated clinical stainer Leica BondMax (Leica Biosystems, Australia). Whole slide fIHC and brightfield differential interference contrast (DIC) images were acquired concurrently with Axioscan 7 Slide Scanner (Zeiss, Germany). Cell segmentation and guided thresholding of CD3+ cells were performed on the fIHC image using HALO (Indica Labs, U.S.A.) to obtain cell centroids and cell-level ground truth labels. Next, each cell was extracted from the DIC image as a 56×56-pixel image patch, yielding 200,000 image patches (equal ratio of CD3+:CD3- cells) that were split 80:20 for training and validation respectively. A held-out dataset consisting of 56,000 image patches was used for testing. The model architecture used is a custom ResNet18 convolutional neural network, and its performance was evaluated with metrics such as F1 score and area under the receiver operating characteristic curve (AUROC). Results: Cell smear slides were successfully stained and imaged, where intact cell morphology was observed across all images. The CD3+ prediction model attained an F1 score of 0.638 and AUROC of 0.87 when applied to the held-out test dataset, indicating that it can differentiate between the unstained brightfield images of CD3+ and CD3- cells. Conclusion: Our cell smear preparation procedure is straightforward and rapid, where the slides could be further processed in an automated staining and imaging pipeline. The acquired images can then be used for identification of protein biomarker expression via IHC. If prediction via DL models is applied, then immunostaining could be omitted. Future works include applying this workflow to all other samples of bodily fluids from patients to examine other disease-related biomarkers, and to build prediction models on other protein biomarkers. Citation Format: Felicia Wee, Willa Wen-You Yim, Jia Meng, Jeffrey Lim, Li Yen Chong, Craig Ryan Joseph, Lee Hong Song, Jabed Iqbal, Mai Chan Lau, Darren Wan-Teck Lim, Joe Yeong. Single cell AI-prediction of protein biomarker on novel whole slide unstained blood smear images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5087.