Search

Accurate and efficient classification and quantification of CD34+ cells are essential for the diagnosis and monitoring of leukemia. Current methods, such as flow cytometry, are complex, time-consuming, and require specialized expertise and equipment. This study proposes a novel approach for the label-free identification of CD34+ cells using a deep learning model and lens-free shadow imaging technology (LSIT). LSIT is a portable and user-friendly technique that eliminates the need for cell staining, enhances accessibility to nonexperts, and reduces the risk of sample degradation. The study involved three phases: sample preparation, dataset generation, and data analysis. Bone marrow and peripheral blood samples were collected from leukemia patients, and mononuclear cells were isolated using Ficoll density gradient centrifugation. The samples were then injected into a cell chip and analyzed using a proprietary LSIT-based device (Cellytics). A robust dataset was generated, and a custom AlexNet deep learning model was meticulously trained to distinguish CD34+ from non-CD34+ cells using the dataset. The model achieved a high accuracy in identifying CD34+ cells from 1929 bone marrow cell images, with training and validation accuracies of 97.3% and 96.2%, respectively. The customized AlexNet model outperformed the Vgg16 and ResNet50 models. It also demonstrated a strong correlation with the standard fluorescence-activated cell sorting (FACS) technique for quantifying CD34+ cells across 13 patient samples, yielding a coefficient of determination of 0.81. Bland–Altman analysis confirmed the model’s reliability, with a mean bias of −2.29 and 95% limits of agreement between 18.49 and −23.07. This deep-learning-powered LSIT offers a groundbreaking approach to detecting CD34+ cells without the need for cell staining, facilitating rapid CD34+ cell classification, even by individuals without prior expertise.

Smartphone-based point-of-care testing (POCT) is rapidly emerging as an alternative to traditional screening and laboratory testing, particularly in resource-limited settings. In this proof-of-concept study, we present a smartphone- and cloud-based artificial intelligence quantitative analysis system (SCAISY) for relative quantification of SARS-CoV-2-specific IgG antibody lateral flow assays that enables rapid evaluation (p = 0.008; Pearson correlation coefficient: 0.56, p = 0.012) between the OD450 values of the enzyme-linked immunosorbent assay and the antibody levels obtained by SCAISY. This study suggests that SCAISY is a simple and powerful tool for real-time public health surveillance, enabling the acceleration of quantifying SARS-CoV-2-specific antibodies generated by either vaccination or infection and tracking of personal immunity levels.

The complete blood count (CBC) is one of the most important clinical steps in clinical diagnosis. The instruments used for CBC are usually expensive and bulky and require well-trained operators. Therefore, it is difficult for medical institutions below the tertiary level to provide these instruments, especially in underprivileged countries. Several reported on-chip blood cell tests are still in their infancy and do not deviate from conventional microscopic or impedance measurement methods. In this study, we (i) combined magnetically activated cell sorting and the differential density method to develop a method to selectively isolate three types of leukocytes from blood and obtain samples with high purity and concentration for portable leukocyte classification using the lens-free shadow imaging technique (LSIT), and (ii) established several shadow parameters to identify the type of leukocytes in a complete leukocyte shadow image by shadow image analysis. The purity of the separated leukocytes was confirmed by flow cytometry. Several shadow parameters such as the “order ratio” and “minimum ratio” were developed to classify the three types of leukocytes. A shadow image library corresponding to each type of leukocyte was created from the tested samples. Compared with clinical reference data, a correlation index of 0.98 was obtained with an average error of 6% and a confidence level of 95%. This technique offers great potential for biological, pharmaceutical, environmental, and clinical applications, especially where point-of-care detection of rare cells is required.