Artificial Intelligence Advances Lymphoma Diagnosis, Biomarker Identification, and Precision Treatment

Clinical Summary:

• Wenchao Han, PhD, examines the emerging role of artificial intelligence (AI) in lymphoma diagnosis and management, including histopathologic classification, multimodal data integration, biomarker identification, and prediction of treatment response. 
• AI-based approaches may enhance recognition of subtle morphologic patterns, support classification of diagnostically challenging lymphoma cases, and integrate molecular, immunophenotypic, and clinical information to refine assessment and identify potential predictive biomarkers. 
• These developments highlight the potential for AI to support more individualized lymphoma care while emphasizing the need for more rigorous external validation, ongoing performance monitoring, multidisciplinary collaboration, and continued physician oversight. 

Wenchao Han, PhD, Mayo Clinic, Rochester, Minnesota, discusses the emerging role of artificial intelligence (AI) in lymphoma diagnosis and management, including applications in diagnostic classification, biomarker discovery, and treatment response prediction. 

Dr Han reviews how AI-based imaging analysis and multimodal data integration may improve the evaluation of challenging lymphoma cases by combining morphologic, immunophenotypic, molecular, and clinical information. He also discusses the potential for AI to refine prognostic and predictive biomarker identification, highlighting the importance of external validation, continuous performance monitoring, and multidisciplinary oversight before these technologies can be reliably integrated into clinical practice.

Transcript:

Thank you for the opportunity to discuss the emerging role of AI in lymphoma diagnosis and the management at LL&M Congress. 

My name is Wenchao Han and I am with the department of lab medicine and pathology at Mayo Clinic. I'm a faculty member in the clinical consultant in computational pathology and informatics, and I serve as a technical director of AI operations in the Blood Bites Research AI Laboratory. 

I'd like to focus on 3 main areas: First, how AI may support a lymphoma diagnosis and the subtype classification, particularly in challenging cases, second, the emerging role in biomarker identification and the prediction and prognosis in the treatment and response, and finally, some important considerations as we think about bringing this technology into clinical practice.

Lymphoma is an area where diagnosis can be particularly complex, and we are increasingly integrating information from morphology, immunophenotype, flow cytometry, cytogenetics, molecular testing, and clinical findings. 

In many cases, the diagnosis is straightforward, but there are also challenges or ambiguous cases where morphology overlaps, immunophenotype is atypical, and the tissue is limited or the findings do not fit neatly into a single diagnosis category– this is an area where I think AI can have significant potential.

AI-based imaging analysis may help us identify subtle morphology patterns that are difficult to quantify consistently with a visual assessment alone. For example, quantify cell populations or recognize the complex patterns across the host line imagery of tissue samples. 

Potentially even more powerful is the ability to integrate multiple data modalities. An AI system could combine morphology with immunophenotypic molecular clinical information and help identify which diagnostic probabilities are mostly consistent with the overall pattern. 

In a difficult case, I see these less, as AI gives us the diagnosis and provide physician support. It may highlight an unusual feature, suggest a difficult diagnosis, identify cases that deserve additional testing, and tell us that a case is sufficiently different from a training data set that a greater caution is warranted.

The second area is the biomarker prognosis and the prediction of the treatment and response. Traditionally, many biomarkers are based on individual measurements. For example, a particular mutation, protein expression level, or genetic alteration but, the lymphoma biology is much more complex than any single variable. AI gives us the opportunity to analyze large numbers of variables simultaneously. Some of these variables may not even be obvious to us visually. 

Ultimately, the goal is not simply to predict prognosis. The more clinical meaningful question is whether we can predict which patient is most likely to benefit from particular a therapy or treatment. This is a much more difficult problem, and I think we need to be careful about distinguishing prognostic biomarkers from truly predictive biomarkers. But this is also where AI could potentially contribute to precision medicine by helping us move from broader disease categories towards more individualized treatment decisions. 

Finally, clinical implementation is critical. AI can perform extremely well in research dataset and still fail when introduced into a different clinical environment. Differences in patient populations, specimen preparation, staining, scanners, laboratory workflows, and clinical practice can all affect the performance. 

Therefore, validation cannot be an afterthought. We need robust validation ideally across institutions and the diverse patient populations. We need to understand not only overall accuracy, but also where and why a model fails. 

We also need ongoing monitoring after implementation because clinical practice technology and patient populations change over time. Perhaps most importantly, successful implementation requires close collaboration between pathologists, clinicians, data scientists, and the AI developers. 

At the same time, we need to maintain appropriate skepticism, rigorous validation, regulatory and quality oversight, and ultimately human responsibility for clinical decision making. 

I think the future of AI in lymphoma is about augmenting that judgment where additional information can have the greatest value for patients.