Instant Insights: Bringing AI Pathology to Any Lab

Author: Denis Avetisyan

A new software platform unlocks real-time computational pathology analysis on standard hardware, democratizing access to AI-powered diagnostics.

OnSight Pathology is a platform-agnostic system enabling real-time, AI-driven analysis of histopathology images without requiring specialized infrastructure.

Despite advances in artificial intelligence, the widespread adoption of computational pathology remains hindered by software compatibility issues and the need for specialized hardware. Here, we present OnSight Pathology: A real-time platform-agnostic computational pathology companion for histopathology, a novel software solution enabling real-time AI-powered analysis of digital pathology images directly on consumer-grade computers. This platform-agnostic approach removes key barriers to deployment, facilitating cost-effective and secure integration into existing research and clinical workflows. Could this accessible, real-time AI assistance fundamentally reshape the future of histopathological analysis and broaden access to expert-level diagnostics?

The Inevitable Bottleneck: Why Eyes Aren’t Enough

Pathological diagnosis, for much of medical history, has fundamentally depended on a skilled pathologist’s interpretation of visual cues from tissue samples. While expertise remains crucial, this reliance on subjective assessment introduces inherent variability – different pathologists, even when examining the same specimen, may arrive at differing conclusions. This inter-observer variation isn’t necessarily due to incompetence, but rather the subtle nuances within biological tissues that lie at the edge of human perceptual ability. Factors like staining inconsistencies, the inherent complexity of cellular morphology, and even daily fluctuations in a pathologist’s focus can all contribute to diagnostic discrepancies. Consequently, a significant portion of pathology cases require a second opinion or further investigation, increasing healthcare costs and potentially delaying appropriate patient care. The field is actively seeking methods to standardize visual interpretation and minimize this subjectivity, recognizing that even minor improvements in diagnostic accuracy can have a substantial impact on patient outcomes.

The escalating demands on modern pathology departments, driven by aging populations and increasingly complex disease presentations, necessitate a shift towards more efficient and objective diagnostic methods. Pathologists are facing ever-growing caseloads, requiring them to analyze a greater volume of samples with decreasing turnaround times. This pressure, coupled with the subtleties inherent in many disease diagnoses, increases the potential for errors and delays in patient care. Consequently, there is a critical need for tools that can augment a pathologist’s expertise – technologies like digital pathology, artificial intelligence-assisted image analysis, and advanced molecular diagnostics – which offer the promise of improved accuracy, reproducibility, and throughput, ultimately supporting clinicians in delivering timely and effective treatment.

The demand for rapid intraoperative pathology assessments, particularly during procedures like frozen section analysis, presents a significant challenge to current diagnostic workflows. Pathologists are frequently tasked with providing immediate, definitive diagnoses to guide surgical decision-making, often within minutes. However, traditional methods can be time-consuming, relying on manual tissue processing, staining, and microscopic evaluation – steps prone to delays and subjective interpretation. This creates a critical bottleneck, potentially prolonging surgeries, increasing patient anxiety, and, in some cases, necessitating further procedures if initial assessments prove inaccurate. Consequently, there is growing interest in technologies that can accelerate these processes, offering real-time or near real-time analysis to enhance surgical precision and improve patient outcomes.

OnSight Pathology: A Platform, Not a Panacea

OnSight Pathology is engineered as a platform-agnostic software solution, meaning it is designed to function independently of specific digital pathology platforms or hardware configurations. This is achieved through compatibility with standard image formats and the utilization of readily available screen capture technology, enabling integration into existing laboratory workflows without requiring substantial infrastructure changes or proprietary system dependencies. The software’s modular design allows it to interface with a variety of scanning devices and image storage systems, facilitating a streamlined implementation process and minimizing disruption to current pathology practices. This approach maximizes accessibility and scalability for laboratories of varying sizes and technological setups.

OnSight Pathology utilizes deep learning models to automate and standardize quantitative pathology tasks. Specifically, the platform employs these models for tumor classification, enabling automated identification of cancer subtypes; mitosis detection, which quantifies cellular proliferation for grading and prognosis; and Ki-67 quantification, assessing the percentage of actively dividing cells within a tissue sample. These automated analyses provide objective, reproducible data, reducing inter-observer variability and improving diagnostic accuracy compared to manual assessment techniques.

OnSight Pathology digitizes pathology slides for analysis using standard screen capture technology, supplemented by hardware such as the OpenOcularAdapter for improved image acquisition. This digitization process enables the application of deep learning algorithms for tasks like tumor classification. Performance metrics demonstrate a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.99 when the system classifies tumors within The Cancer Genome Atlas (TCGA) cohorts, indicating a high degree of accuracy in distinguishing between cancerous and non-cancerous tissues.

Under the Hood: Calibration and the Illusion of Objectivity

Accurate data calibration within the OnSight Pathology platform is fundamental to generating reliable quantitative results. This process involves correcting for systematic errors and variations introduced by imaging hardware, staining protocols, and tissue preparation. Calibration ensures that pixel intensities consistently correlate with biological signals, allowing for precise and reproducible measurements of features like cell counts, protein expression levels, and morphological characteristics. Without rigorous calibration, quantitative analysis is susceptible to inaccuracies that can impact diagnostic and research outcomes; the platform employs standardized calibration procedures and quality control metrics to minimize these errors and maintain data integrity across different samples and experiments.

Image segmentation within OnSight Pathology utilizes algorithms to partition digital pathology images into multiple regions, enabling the isolation and quantification of specific biological features. This process identifies areas of interest, such as cells or tissue structures, and delineates their boundaries with pixel-level precision. For example, Ki-67 expression, a marker of cellular proliferation, is quantified by segmenting positively stained nuclei. The system calculates metrics like the percentage of Ki-67 positive cells within a defined region of interest, providing a quantitative assessment of proliferative activity. Accurate segmentation is achieved through a combination of color deconvolution, morphological operations, and machine learning-based approaches, allowing for objective and reproducible analysis.

OnSight Pathology integrates deep learning models, specifically Vision Transformers (ViT) and You Only Look Once (YOLO), to enhance its analytical capabilities beyond traditional image analysis. These models facilitate precise object detection and classification within pathology images, automating tasks such as tumor classification, mitosis detection, and MIB (Ki-67) segmentation. Performance benchmarks demonstrate inference times of less than 400 milliseconds for these analytical processes when executed on standard consumer-grade laptop hardware, indicating a capacity for real-time or near real-time analysis without requiring specialized high-performance computing infrastructure.

From Pixels to Prose: Automating the Narrative, Not the Expertise

OnSight Pathology leverages the capabilities of the Lingshu multimodal language model to transform complex histological images into detailed textual reports. This innovative approach allows the system to automatically generate descriptions of cellular structures, tissue patterns, and other key features observed within a sample. By effectively ‘reading’ the visual information contained in the image, Lingshu constructs a narrative that can be used to support diagnostic assessments and facilitate communication between pathologists. The integration of this AI-powered description generation not only streamlines the reporting process but also ensures consistency and accuracy in documenting microscopic findings, potentially reducing interpretive variability and improving patient care.

The integration of multimodal AI into pathology workflows dramatically enhances the dissemination of critical diagnostic information. By automatically generating detailed textual reports from histological images, the technology transcends traditional static images, offering a richer, more accessible record for both immediate interpretation and long-term reference. This capability fosters seamless knowledge sharing among pathologists, enabling efficient collaboration on complex cases and facilitating consistent, standardized reporting across institutions. The resulting comprehensive documentation not only improves diagnostic accuracy but also serves as a valuable resource for educational purposes and the advancement of pathological research, ultimately streamlining workflows and improving patient outcomes through enhanced communication and collaborative expertise.

OnSight Pathology dramatically expands access to specialized diagnostic expertise through a robust telepathology capability. The platform allows pathologists to remotely consult on cases and deliver diagnoses from any location with internet connectivity, overcoming geographical barriers to care. This is particularly impactful in areas with limited access to subspecialty expertise, or during times requiring rapid second opinions. Crucially, the system demonstrates exceptional performance in critical applications; specifically, it achieves 99.98% accuracy in classifying metastatic tumors using frozen sections – a speed and precision vital for timely cancer staging and treatment planning. This level of accuracy, combined with remote accessibility, positions OnSight Pathology as a transformative tool in modern pathology practice.

Future Directions: Scaling AI, Accepting Limitations

The foundation of OnSight Pathology’s reliability lies in the extensive training and validation of its deep learning models using comprehensive datasets, notably TCGAData-a resource containing genomic and clinical information from thousands of cancer patients. This rigorous process isn’t merely about quantity; it’s about exposing the algorithms to the immense variability inherent in cancer pathology, encompassing diverse tissue types, grades, and stages. By learning from such a broad spectrum of cases, the models develop a heightened ability to generalize and maintain accuracy when encountering previously unseen data. This focus on robustness is critical for clinical application, where consistent and dependable performance is paramount; it minimizes the risk of false positives or negatives, ensuring that diagnostic and treatment decisions are guided by trustworthy insights derived from the digital pathology analysis.

A defining characteristic of this platform is its capacity for real-time inference, enabling exceptionally swift analysis of pathology slides. This immediate processing is not merely a technical achievement, but a crucial step towards transforming clinical workflows; it allows pathologists to receive data-driven insights during the diagnostic process, rather than awaiting results from lengthy, traditional analyses. The speed facilitates more informed, on-the-spot decision-making, potentially accelerating treatment planning and improving patient care timelines. Such rapid assessment is particularly vital in aggressive cancers where prompt intervention significantly impacts prognosis, and offers a pathway towards a more proactive and personalized approach to oncology.

OnSight Pathology fundamentally alters the landscape of cancer diagnosis by integrating artificial intelligence into established pathology workflows. The platform’s ability to deliver objective, quantitative data – demonstrated by a ROC AUC of 0.75 on the ICPR 2014 mitosis detection challenge – moves beyond subjective human assessment, offering a consistent and reliable foundation for treatment planning. This streamlined approach not only accelerates the diagnostic process but also unlocks the potential for truly personalized cancer care, tailoring therapies to the unique characteristics of each patient’s disease and ultimately improving outcomes through more precise and informed clinical decisions.

The pursuit of seamless integration, as presented with OnSight Pathology’s platform-agnostic design, invariably courts future complications. It strives to bridge the gap between research and readily available hardware, a noble goal, yet one destined to encounter the realities of diverging standards and unforeseen compatibility issues. As Yann LeCun observes, “If you want to be a systems builder, you need to be comfortable with the fact that 90% of your time will be spent debugging other people’s code.” This is especially true when aiming for broad accessibility; the elegance of real-time inference on consumer hardware will, inevitably, be tested by the chaotic variance of production environments. Tests are a form of faith, not certainty, and OnSight’s success will be measured not by initial performance, but by its resilience against the inevitable entropy of real-world deployment.

What’s Next?

The promise of widespread, accessible computational pathology hinges on removing dependencies. This work addresses that, shifting processing to the edge. But the edge, predictably, introduces new surfaces for failure. The bug tracker will soon hold tales of driver incompatibilities, unexpected hardware acceleration quirks, and the eternal struggle to support ‘just one more’ image format. The platform-agnostic approach isn’t a triumph over constraint, it’s a deferral-a broadening of the attack surface.

Real-time inference is compelling, but clinical workflows aren’t defined by speed. They’re defined by exceptions. The system will perform flawlessly on the curated datasets, of course. It’s the rare, the unexpected, the poorly prepared slide that will expose the limits of any algorithm, and the fragility of a system designed for convenience. The real test isn’t accuracy on benchmarks, it’s resilience in the face of chaos.

The next iteration won’t be about better algorithms, or faster processing. It will be about graceful degradation. About acknowledging that perfect prediction is a mirage, and building systems that admit error, and still provide value. The aim isn’t to solve pathology, it’s to provide tools that amplify the expertise of those who do. They don’t deploy – they let go.

Original article: https://arxiv.org/pdf/2512.04187.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/