New Prognostic Biomarker for Hodgkin Lymphoma Identified Using AI

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a recent study, researchers at Justus Liebig University and Heidelberg University used deep learning to quantify a potential new biomarker — the amount of weakly stained collagen fibers in biopsies — and correlated this with patient outcomes.1 Their aim was to evaluate whether AI-aided image analysis could enable accurate quantification of histological features predictive of prognosis in Hodgkin lymphoma.

The study showed that quantification of collagen fibers using picrosirius red could discriminate cases based on the risk of progression and that the percentage of weakly stained fibers was significantly associated with treatment response.

“Our results suggest that patients with poor treatment response may have more aberrant fibers with weak staining and irregular consistency compared to patients with good response,” said Dr. Ila Motmaen, the first author of the study.

He cautioned, however, that the prognostic power of aberrant fibers is not powerful enough to be used as a sole biomarker.

The report was published in the European Journal of Haematology.

Study Rationale: Leveraging AI to Identify Novel Biomarkers for Hodgkin Lymphoma

Hodgkin lymphoma, a cancer of the lymphatic system, is common in children and adolescents and affects over 80,000 people worldwide annually. Although the disease is cured in most cases, 10%–15% of patients experience recurrence after first-line chemotherapy or show resistance to conventional therapies.1

Currently, there are no reliable biomarkers to identify high-risk patients early during treatment. Although quantification of MMP9-positive cells in biopsies has been suggested as a predictor of treatment outcomes in patients with Hodgkin lymphoma, manual quantification is subjective and labor intensive.

In this study, researchers leveraged AI to analyze tumor biopsies and identify a candidate prognostic biomarker for Hodgkin lymphoma. Dr. Motmaen noted that accurately identifying high-risk patients early on could better inform treatment decisions. More aggressive therapy or experimental regimens could be administered to patients at risk of recurrence, whereas low-risk patients could be spared the unnecessary toxicity of intensive regimens.

Collagen Fibers as a Prognostic Indicator

Previous studies have shown that the collagen scaffolding surrounding lymphoma cells plays a role in tumor progression and invasion. Matrix metalloproteinases (MMPs) are enzymes that break down collagen fibers, and elevated MMP levels are associated with poor prognosis in patients with lymphoma.

This motivated researchers to explore whether visual features of collagen in biopsies, such as the degree of collagen fiber staining, might constitute a quantitative prognostic biomarker. They hypothesized that image analysis for collagen fibers could identify patients with abnormal collagen patterns reflecting increased MMP activity and poor prognosis. Manual scoring of such subtle visual features is impractical for routine clinical use. Therefore, researchers used AI to detect and quantify collagen fibers in whole slide images (WSI) of routine biopsies.

 Approach: Applying Deep Learning to Pathology Images

Researchers used a convolutional neural network (CNN) called YOLOv4 to analyze digitized pathology slides from Hodgkin lymphoma biopsies. Eighty-three cases were selected, representing a range of treatment responses based on post-chemotherapy qualitative positron emission tomography (qPET) imaging.1 PET response is correlated with prognosis, with minimal residual lymphoma activity indicating a favorable prognosis.

Biopsies were stained using picrosirius red dye, which stains collagen fibers. The stained slides were digitized using whole slide imaging, and the AI model was trained to recognize and quantify weakly stained and strongly stained collagen fibers. Algorithm training involved manual annotation of collagen fibers in the training dataset comprising 30 WSIs.

Subsequently, the trained model was used to analyze the remaining 53 WSIs and quantify the percentage of weakly stained fibers in each case. This percentage was correlated with prognosis based on qPET scores.

For validation, a subset of slides was also stained and analyzed for MMP9 expression using the same approach. Dr. Motmaen explained that MMP9 degrades collagen and that MMP9 levels have been linked to lymphoma prognosis.

He added that the degree of abnormal collagen patterns (i.e., percentage of weakly stained fibers) in each WSI was compared to the percentage of MMP9-positive (abnormal) cells to verify results; the percentage of MMP9-negative cells was matched to the percentage of strongly stained collagen fibers.

Prognostic Value of Collagen Fibers in Hodgkin Lymphoma

The study revealed a significant association between higher percentages of weakly stained collagen fibers and worse post-chemotherapy qPET scores, indicating a higher risk of recurrence after treatment (P = 0.0185).1

Cases with unfavorable qPET scores (suggesting a higher risk of recurrence) had an average of 18% weakly stained fibers. In contrast, lower-risk cases with favorable qPET responses had 10%–14% weakly stained fibers.

The AI model achieved reasonable accuracy in discriminating low- from high-risk cases based on the degree of weakly stained collagen fibers, providing an area under the curve (AUC) of  0.79.

Analysis of MMP9-stained slides from the same patients revealed that high MMP9 expression was associated with a higher percentage of abnormal, weakly stained collagen patterns.

“Given the role of MMP9 in collagen metabolism and its association with lymphoma prognosis, this finding strengthens the biological rationale for using weakly stained fibers as a prognostic marker,” said Dr. Motmaen.

Collectively, these results suggest that the degree of weak collagen fiber staining could be a quantitative biomarker for predicting the prognosis of Hodgkin lymphoma. However, the relationship between the percentage of weakly stained collagen fibers and prognosis was not strong enough for clinical use.

Future Work

“While promising, the prognostic value of quantified weakly stained fibers was not strong enough for high confidence prognosis,” Dr. Motmaen noted. The prognostic value of collagen fibers in Hodgkin Lymphoma warrants further validation in larger patient cohorts.

Improving the model to detect collagen fibers more consistently may enhance its clinical utility, and using additional annotated training data could reduce errors in distinguishing weak from strong staining. Furthermore, incorporating additional biomarkers could improve the prognostic performance of the model.

Even though the prognostic model needs to be improved, this study demonstrates the feasibility of using AI to identify new prognostic biomarkers in lymphoma. The findings lay the groundwork for developing AI-based tools to automate prognostication based on standard biopsy slides and guide personalized treatment in Hodgkin lymphoma.


  1. Motmaen I, Sereda S, Brobeil A, et al. Deep-learning based classification of a tumor marker for prognosis on Hodgkin’s disease. Eur J Haematol. August 2023:1–7. doi:10.1111/ejh.14066

Share This Post

Leave a Reply