Despite rapid progress in digital pathology, many AI models remain confined to narrow, single task applications that struggle to translate to clinical applications. And for the diagnosis of oral epithelial dysplasia (OED), invasive biopsy and histopathology remain the procedure of choice.

Here, we speak with Adam Shephard, Assistant Professor at the University of Warwick, who outlines an alternative approach to OED diagnostics utilizing AI applications and improving outcomes for patients.

What first led you to study OED using AI?

I completed my PhD in neuroimaging, using MRI to study conditions including pediatric brain injury. Through this work, I became increasingly interested in understanding disease at a much finer, microscopic level.

This led me to computational pathology and, in particular, OED. I joined the Tissue Image Analytics (TIA) Centre at the University of Warwick, working with Professor Nasir Rajpoot, and began researching OED as part of a Cancer Research UK–funded project in collaboration with Ali Khurram and Hanya Mahmood at the University of Sheffield.

During this work, I gained a deeper appreciation of this understudied precancerous condition. OED is relatively common, particularly in lower- and middle-income countries, and a proportion of cases progress to oral cancer. One of the main challenges is that the current gold standard for risk assessment – histological grading – is subjective and has limited ability to predict which lesions will become malignant.

This makes OED a compelling area for computational pathology and AI, with the potential to improve risk stratification and support more informed clinical decision-making.

What are the main problems pathologists face when grading OED, and how can this affect patient care?

Grading OED is challenging for several reasons. First, there can be significant variation between pathologists, meaning the same lesion may receive different grades from different observers. Second, grading requires assessment of multiple subtle morphological changes throughout the epithelium, making it a complex and subjective process.

A further challenge is that histological grade does not always predict clinical outcome. Some lesions classified as mild may progress to cancer, while some severe lesions remain stable. As a result, treatment decisions based solely on grading can be difficult and may increase the risk of over- or under-treatment.

These limitations highlight the need for more objective approaches to risk assessment that can better predict disease progression and support clinical decision-making.

What is HoVer-Net+ and how was it used in your research?

HoVer-Net+ is an AI model we developed to analyze histology images at the nuclear level. It builds on earlier work from the TIA Centre by identifying individual cells and recognizing different tissue regions within a sample. We adapted and trained the model specifically for OED.

This approach allows us to extract biologically meaningful features that are closer to the types of changes a pathologist may assess, rather than relying only on “black box” predictions. This may help make the model’s outputs more interpretable and clinically acceptable.

These extracted features were then used in later models to predict cancer risk in OED. They also helped identify potentially important biomarkers, including increased lymphocyte infiltration associated with malignant progression.

Your work looked at predicting which OED cases may become cancerous. How important is this for improving early detection and treatment?

Predicting which cases of OED are most likely to progress to cancer is central to improving patient care. More accurate risk assessment could help clinicians identify high-risk patients earlier and guide more appropriate intervention.

Our findings suggest that AI-based models may provide more consistent risk predictions than histological grading alone. Importantly, these tools are not intended to replace pathologists, but to support them by providing additional information that can be incorporated into clinical decision-making.

In practice, this could involve providing a risk score alongside visual indicators highlighting tissue regions that contributed to the prediction. Such tools could support faster, more objective, and more consistent assessments while maintaining the pathologist's role in interpretation and diagnosis.

Ultimately, improved risk stratification could help ensure that patients at greatest risk of malignant progression receive timely treatment and follow-up, which is critical for achieving better outcomes.

What were the main findings of the study, particularly in terms of new biomarkers or features linked to cancer risk?

A key outcome of our work has been the development of quantitative risk assessment tools, including the OMTscore and, more recently, the ODYN model. These models predict the risk of malignant progression directly from tissue images and, in some settings, have achieved performance comparable to that of expert pathologists.

An important advantage of using interpretable image features is that they can also provide biological insights. For example, we identified associations between cancer progression and the presence of peri- and intra-epithelial lymphocytes. This was a particularly interesting finding because it differs from the traditional view of lymphocytes as being solely protective in cancer.

These insights not only improve risk prediction, but also contribute to a better understanding of the biological processes underlying disease progression.

How did these AI-based biomarkers perform compared with traditional methods of grading OED?

Overall, the AI-derived biomarkers performed well and showed strong agreement with expert pathologists on internal datasets. Performance declined slightly when the models were tested on external datasets, a challenge commonly seen in pathology because of differences in staining protocols, image quality, patient populations, and clinical practice between institutions.

Despite these variations, the models were still able to meaningfully stratify patients according to their risk of malignant progression. This suggests that they are capturing biologically relevant features that can generalize beyond a single center.

Why is it useful to combine histology with clinical data in multimodal AI models, rather than relying on slide images alone?

Combining histology with clinical data can provide a more complete picture of the patient. While tissue images capture detailed biological information, clinical factors such as medical history, smoking status, and alcohol consumption provide important context. In OED, for example, smoking and alcohol use are known risk factors for malignant progression.

We have not yet done extensive work combining these data types in OED, largely because collecting consistent clinical data across multiple sites can be challenging in retrospective studies. However, in other projects, we have found that histological and clinical data can each provide independent predictive value, and that combining them can improve risk assessment.

The benefits of multimodal models depend on the information available. In some cases, clinical variables may already capture much of the relevant pathological information, limiting the added value of imaging data. Understanding how different data sources complement one another remains an important area of ongoing research.

Looking ahead, how do you see AI tools like these changing pathology practice and improving outcomes for patients with OED?

I expect AI to become an increasingly integrated part of routine pathology practice. As more laboratories adopt digital pathology workflows, larger datasets will become available, enabling AI models to become more robust, reliable, and broadly applicable. Initiatives such as Secure Data Environments in the UK are helping researchers access and analyze these data securely.

There is also growing interest in multimodal approaches that combine histology, clinical, molecular, and other data types to provide a more comprehensive understanding of disease progression. Emerging agentic AI systems may further enhance this by supporting reasoning and the use of specialized analytical tools.

For conditions such as OED, where risk stratification remains challenging, AI has the potential to support more objective and reproducible assessments. With appropriate validation, regulation, and clinical oversight, these tools could help identify high-risk patients earlier, enabling more timely intervention and improved outcomes.

Newsletters

Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.

I have read and understand the Privacy Notice*

Subscribe