Cytopathology, particularly fine needle aspiration (FNA) cytology, relies heavily on optical microscopy for cancer diagnosis. However, current methods face significant limitations. Conventional microscopes offer a trade-off between field-of-view (FOV) and resolution; achieving both large FOVs necessary to cover the entire cytology smear and high resolution needed for cellular-level detail is challenging. The thickness of FNA smears (often >50 µm) further complicates the process, demanding 3D imaging capabilities. While whole slide imaging (WSI) systems exist, they are slow and expensive, often taking over an hour per slide, hindering rapid on-site evaluation (ROSE) and delaying diagnosis. This slow speed, coupled with high cost, limits the widespread clinical adoption of WSI for cytopathology. Alternative rapid imaging techniques like Structured Illumination Microscopy (SIM) and Fourier Ptychography offer improved speed and resolution but still fall short in FOV compared to the size of a typical cytology slide and struggle with thicker specimens. The need for a faster, more cost-effective, and high-resolution system for 3D imaging of thick cytology smears is crucial for improving the speed and accuracy of cancer diagnosis. The authors address this challenge by developing a novel multi-camera array scanner (MCAS).

The paper reviews existing limitations in optical microscopy for cytopathology, highlighting the trade-off between FOV and resolution. The slow speed and high cost of current WSI systems, particularly for thick cytology smears, are discussed, along with the limitations of alternative rapid imaging techniques like SIM and Fourier Ptychography in terms of FOV and handling of thicker specimens. The existing literature underscores the need for a system that can achieve high resolution, large FOV, and rapid 3D imaging capabilities for effective digitization of thick cytology smears. The authors also review existing machine learning techniques used for the analysis of cytology images, setting the stage for their own implementation of such techniques with the data obtained using their new MCAS system.

The core of the methodology is the development and application of the Multi-Camera Array Scanner (MCAS). The MCAS consists of an array of 48 individual 13-megapixel CMOS image sensors, each equipped with a custom-designed objective lens. This array allows for simultaneous imaging of different areas of the sample, significantly parallelizing the imaging process. Two different custom-designed lenses were used, one with 0.3 NA (1.2 µm resolution) and another with 0.5 NA (0.6 µm resolution), providing a comparison between speed and resolution. The sensors are arranged in a 6x8 grid with a 9 mm spacing, minimizing the required scanning distance. Three-axis motorized stages facilitate precise specimen movement for both lateral and axial scanning. The system captures 0.63 gigapixels per snapshot and transfers data through a field-programmable gate array (FPGA) at up to approximately 5 gigapixels per second. For 3D imaging, the sample is scanned across multiple axial slices. The authors detail the image processing pipeline, including stitching algorithms to combine images from individual cameras and a contrast metric for identifying the sharpest focus plane within z-stacks. A custom viewer, Gigaviewer, is used for visualizing the resulting 16-gigapixel (or 80 gigapixel for 5 axial planes) images. The machine learning methodology involves two key approaches: 1) object detection using YOLOv7 for adenocarcinoma localization in lung specimens, and 2) slide-level classification using Multiple Instance Learning (MIL) for classifying slides as adenocarcinoma-positive or negative. The data for these machine learning tasks consisted of Diff-Quik stained cytology samples from archival cases, including both adenocarcinoma-positive and benign samples. For the object detection task, images were divided into smaller patches, labeled by clinicians, and used to train the YOLOv7 model. For the slide-level classification, a MIL approach was used, treating each slide as a "bag" of image patches, leveraging the fact that the presence of at least one positive instance within a bag (slide) indicates a positive class label. 4-fold cross-validation was employed to assess the model's performance with the limited dataset.

The MCAS demonstrated significantly faster 3D imaging of cytology smears compared to conventional WSI systems. Scanning three entire slides in 3D took less than 5 minutes with the 0.3 NA lens and 40 minutes with the 0.5 NA lens. This is a significant improvement over the hour or more typically required by existing WSI systems. The object detection model trained on MCAS images of lung specimens achieved a recall of 0.73 in identifying adenocarcinoma regions. The slide-level classification model, utilizing MIL, demonstrated high accuracy (0.930), precision (0.931), recall (0.938), and an area under the ROC curve (AUC) of 0.969, effectively classifying slides as adenocarcinoma-positive or negative. In the adequacy assessment task using thyroid specimens, the MCAS-based ROSE decisions achieved 100% sensitivity and 94.4% specificity compared to standard microscopic evaluation.

The findings address the critical need for rapid and high-resolution 3D imaging in cytopathology. The MCAS significantly accelerates the digitization of cytology smears, facilitating rapid on-site evaluation and potentially reducing diagnostic delays. The integration of machine learning algorithms further enhances the efficiency of the workflow, assisting pathologists in identifying key regions of interest (adenocarcinoma in the lung specimens) and classifying slides at a slide level. The high performance of both the object detection and slide-level classification models highlights the potential of MCAS for improving the accuracy and efficiency of cytopathology. The results suggest that MCAS can play a crucial role in optimizing cytopathology workflows, particularly in high-throughput settings or situations requiring rapid diagnosis.

The MCAS offers a significant advancement in cytopathology by enabling rapid and high-resolution 3D imaging of thick cytology smears. Its speed and parallelized approach vastly outperform conventional WSI systems. The integration of machine learning further automates analysis, potentially reducing workload and improving diagnostic consistency. Future research could focus on further cost reduction through mass production of components, exploring different illumination strategies (e.g., variable angle illumination), utilizing the full 3D dataset for machine learning, and incorporating fluorescence capabilities for enhanced diagnostic information and the possibility of digitizing the staining process.

The study used a relatively small dataset for training the machine learning models, which could limit the generalizability of the results. The performance of the object detection model could be improved by addressing inconsistent labeling and potentially by increasing the dataset. The assumption made in the MIL model that at least one positive instance exists in a positive slide may not always hold true. While the MCAS demonstrates significant speed improvements, the acquisition time with the higher-resolution 0.5 NA lens is still considerable.