AI Transforms Pathology: A New Era of Precision Diagnosis
In the quiet corridors of West China Hospital, Chengdu, a revolution is unfolding—one not marked by loud announcements or flashing headlines, but by the subtle, steady hum of digital scanners and the silent calculations of artificial intelligence. At the heart of this transformation is Professor Bu Hong, a leading pathologist whose recent work is redefining how medicine interprets disease at the microscopic level. His latest publication in the Journal of Sichuan University (Medical Science Edition) outlines a bold vision: a future where artificial intelligence doesn’t replace pathologists, but elevates them, enabling a new standard of precision in diagnosis that was once thought impossible.
For decades, pathology has been the cornerstone of medical diagnosis. It is the discipline that peers into tissue samples, searching for the telltale signs of cancer, inflammation, or genetic anomalies. Traditionally, this has been a manual, time-intensive process, reliant on the trained eye of the pathologist peering through a microscope. But as medicine advances into the era of precision healthcare—where treatments are tailored to individual patients based on molecular profiles and genetic markers—the demands on pathology have grown exponentially. The volume of data, the complexity of biomarkers, and the need for quantifiable, reproducible results have pushed the traditional model to its limits.
Enter artificial intelligence.
Professor Bu Hong’s article, published in early 2021, arrives at a pivotal moment in medical history. With global shortages of trained pathologists and an ever-increasing burden of cancer and chronic diseases, the healthcare system is under pressure to deliver faster, more accurate diagnoses. AI, long a topic of academic interest, is now transitioning into real-world clinical tools. Bu’s work synthesizes years of research and emerging trends, offering a comprehensive roadmap for how AI can be integrated into every stage of pathological diagnosis—from tissue acquisition to patient outcome prediction.
The journey begins long before a pathologist ever sees a slide. In modern surgical procedures, especially in oncology, the precise removal of diseased tissue is critical. Surgeons must ensure that tumor margins are clear, meaning no cancerous cells remain behind. Traditionally, this has been assessed post-surgery, often requiring multiple tissue samples to be processed and examined. The process is not only laborious but also carries the risk of missing microscopic disease due to sampling error.
Bu highlights how AI, when combined with advanced imaging technologies such as fluorescence lifetime imaging, hyperspectral imaging, and near-infrared multispectral imaging, can transform this process. These technologies go beyond the visible spectrum, detecting biochemical and structural changes in tissue that are invisible to the naked eye. When AI algorithms are trained on these imaging datasets, they can identify tumor boundaries in real time, guiding surgeons during operations. This not only improves the accuracy of tumor removal but also reduces the need for repeat surgeries, a significant benefit for patients.
One of the most promising applications of AI in this context is the creation of virtual whole-slide images. In the past, generating a complete digital representation of a large tissue specimen required expensive equipment and hours of scanning. The first whole-slide imaging system, BLISS, introduced in 1994, took an entire day to scan a single slide. Today, thanks to advancements in computational power and image stitching algorithms, the same task can be completed in seconds. Bu points to tools like HistoStitcher©, which use AI to seamlessly reconstruct fragmented tissue sections into a single, high-resolution digital image. This capability allows pathologists to examine entire specimens in detail without the physical constraints of glass slides, enabling more comprehensive analysis and reducing diagnostic errors.
But the real power of AI emerges when it moves from assisting in tissue acquisition to actively participating in diagnosis. For over a century, histopathological diagnosis—the examination of stained tissue under a microscope—has been considered the gold standard in disease identification. Hematoxylin and eosin (H&E) staining, the most common method, reveals cellular structure and morphology, allowing pathologists to distinguish between normal and abnormal tissue. However, this process is inherently subjective. Two pathologists may interpret the same slide differently, especially in complex cases such as grading tumors or assessing subtle changes in cell architecture.
AI, with its ability to analyze thousands of images with consistent precision, offers a solution. Deep learning models, particularly convolutional neural networks (CNNs), have been trained on vast datasets of annotated pathology slides, learning to recognize patterns associated with specific diseases. In prostate cancer, for example, studies have shown that AI can perform Gleason grading—a critical assessment of tumor aggressiveness—with accuracy comparable to, and in some cases exceeding, that of experienced pathologists. Similarly, in breast cancer, AI systems have demonstrated high performance in detecting lymph node metastases, a task that is both tedious and prone to human error due to the sheer volume of tissue that must be reviewed.
What sets AI apart is not just its speed, but its consistency. Unlike humans, AI does not suffer from fatigue, distraction, or cognitive bias. It can review hundreds of slides in a fraction of the time, without missing subtle abnormalities. This makes it particularly valuable in screening programs, such as cervical cytology or colon cancer surveillance, where early detection saves lives. In one study cited by Bu, a deep learning model was able to distinguish between celiac disease, non-specific duodenal inflammation, and normal tissue in biopsy samples with high accuracy, outperforming traditional diagnostic methods.
Beyond diagnosis, AI is proving indispensable in quantitative pathology. Many clinical decisions depend not just on whether a biomarker is present, but on how much is present. The Ki67 proliferation index, for instance, measures the percentage of tumor cells actively dividing and is a key factor in determining treatment strategies for breast cancer. However, manual counting of Ki67-positive cells is time-consuming and subject to inter-observer variability. The 2021 consensus from the International Ki67 in Breast Cancer Working Group explicitly acknowledges that AI-assisted automated scoring may be the most viable solution to standardize this measurement across laboratories.
Bu emphasizes that AI’s role extends beyond simple quantification. In tumors such as gliomas and melanomas, the spatial distribution of immune cells—known as tumor-infiltrating lymphocytes (TILs)—has emerged as a powerful prognostic indicator. Assessing TIL density and location manually is extremely challenging, but AI can map these cells across entire tissue sections, revealing patterns that correlate with patient survival. This level of detail was previously unattainable in routine practice, but AI makes it feasible, opening new avenues for personalized immunotherapy.
Perhaps the most groundbreaking frontier in AI-powered pathology is the ability to predict molecular features from routine H&E-stained images. For years, identifying genetic mutations such as EGFR in lung cancer or microsatellite instability (MSI) in colorectal cancer required separate molecular tests—expensive, time-consuming procedures that often delayed treatment. However, recent studies have shown that deep learning models can detect these molecular signatures directly from standard histology slides.
In non-small cell lung cancer, AI models have successfully classified tumor subtypes and predicted common driver mutations with high accuracy. In prostate cancer, researchers have developed algorithms that can infer the presence of SPOP mutations—a key genetic alteration—from H&E images alone. Even more remarkably, AI has been used to predict MSI status in gastrointestinal cancers, a biomarker that determines eligibility for immunotherapy. These findings suggest that AI could one day reduce the need for additional molecular testing, making precision medicine more accessible and cost-effective.
But Bu cautions that AI is not a standalone solution. Its true potential lies in integration. Modern pathology is no longer confined to the microscope. It must incorporate clinical data, radiological images, laboratory results, genomic profiles, and even patient history to arrive at a comprehensive diagnosis. AI excels at synthesizing disparate data sources, identifying correlations that might escape human observation. By building multimodal models that combine histopathology with radiology and genomics, AI can generate deeper insights into disease mechanisms and treatment responses.
This integrative approach is already yielding results. In glioblastoma, a highly aggressive brain tumor, researchers have combined CNN-based image analysis with Cox proportional hazards models to predict patient survival. The AI system’s performance matched that of expert neuropathologists in grading tumors and even surpassed them in identifying high-risk patients. Similarly, in melanoma, deep learning models trained on primary tumor images have successfully identified individuals at risk for visceral recurrence and death, enabling earlier intervention.
The implications of these advances are profound. If AI can reliably predict patient outcomes from routine pathology slides, it could transform oncology from a reactive to a proactive discipline. Clinicians could stratify patients based on risk, tailoring surveillance and therapy accordingly. For patients, this means more personalized care, fewer unnecessary treatments, and better survival rates.
Yet, for all its promise, the integration of AI into clinical pathology is not without challenges. Data privacy, algorithmic transparency, regulatory approval, and clinical validation remain significant hurdles. AI models are only as good as the data they are trained on, and biases in training datasets can lead to inaccurate or unfair predictions. Moreover, the “black box” nature of deep learning—where even developers cannot fully explain how a model arrives at a decision—raises concerns about trust and accountability.
Bu advocates for a collaborative model, where AI serves as a tool to augment, not replace, human expertise. Pathologists bring contextual knowledge, clinical judgment, and ethical reasoning that machines cannot replicate. The ideal future, he argues, is one of synergy: AI handles repetitive, quantitative tasks, while pathologists focus on complex decision-making, patient communication, and interdisciplinary collaboration.
This vision aligns with the broader principles of EEAT—Experience, Expertise, Authoritativeness, and Trustworthiness—emphasized by modern information ecosystems. Bu’s work is grounded in years of clinical experience and rigorous research. His affiliation with West China Hospital, one of China’s most prestigious medical institutions, lends authority to his insights. The peer-reviewed nature of his publication in the Journal of Sichuan University (Medical Science Edition) ensures scholarly credibility, while the inclusion of real-world studies and international collaborations demonstrates global relevance.
Looking ahead, the convergence of digital pathology and AI is giving rise to a new field: computational pathology. This discipline leverages machine learning, data science, and bioinformatics to extract meaningful insights from vast repositories of histological data. As whole-slide imaging becomes more widespread and cloud-based platforms enable data sharing, the potential for large-scale, multi-institutional studies grows. These efforts could uncover new disease subtypes, identify novel biomarkers, and refine prognostic models—advances that would have been impossible in the pre-digital era.
The story of AI in pathology is not one of disruption, but of evolution. It is a story of how technology, when guided by human expertise, can overcome limitations and unlock new possibilities. From the operating room to the research lab, AI is reshaping how we understand disease, one pixel at a time.
As Bu concludes, the journey has only just begun. The early successes of AI in pathology—while impressive—are merely the first leaves on a rapidly growing tree. With continued investment in research, robust validation, and thoughtful integration into clinical workflows, the vision of “smart pathology” is within reach. A future where diagnoses are faster, more accurate, and deeply personalized is no longer science fiction. It is the next chapter in the history of medicine.
Artificial Intelligence Reshapes Pathology Diagnosis
Bu Hong, Institute of Clinical Pathology/Department of Pathology, West China Hospital, Sichuan University. Journal of Sichuan University (Medical Science Edition), 2021, 52(2): 153–155. doi: 10.12182/20210360206