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

Researchers from the Erasmus University Medical Centre have developed an artificial intelligence (AI) framework that identifies and quantifies seven key histological structures in melanoma with over 92% accuracy. The novel approach addresses the pressing challenge of inter-observer variability in melanoma assessment that has limited the predictive power of current staging systems. 

“Traditional assessment of mitotic rate in cutaneous melanoma relies on identifying and counting mitotic figures within a 1 mm² hotspot area selected by the pathologist. This technique is naturally subjective,” noted Antien Mooyaart, MD, PhD, the corresponding author of the study. “Our model addressed this challenge by detecting mitoses across the entire tumor area, standardizing mitotic quantification and removing the subjectivity of hotspot selection.” 

The report was published in Pathology – Research and Practice. 

Limitations in Current Melanoma Assessment

Histopathological assessment remains the gold standard for melanoma diagnosis and staging. The American Joint Committee on Cancer (AJCC) staging system primarily uses Breslow thickness and ulceration status to classify early-stage melanoma, having removed mitotic rate from the 8th edition due to high inter-observer variability. 

This removal came despite clear evidence that mitotic activity correlates with melanoma-specific survival. “Mitotic activity in tumor cells may be an indication for the proliferation and therefore an indication of the tumor progression,” Dr. Mooyaart explained, highlighting the importance of this feature despite its current exclusion from formal staging. 

In addition, although tumor-infiltrating lymphocytes, Breslow density, and hair follicle involvement contribute to melanoma behavior, there are no standardized assessment methods for these features. 

Developing a Sequential AI Framework

The team developed a seven-step progressive framework using convolutional neural networks to analyze whole slide images (WSIs) of cutaneous melanoma. The model was built using 157 melanoma WSIs, with each application trained independently to detect different tissue structures and features including the whole tissue, the tumor microenvironment, hair follicles, sebaceous glands, the epidermis, ulceration, melanoma cell area, and mitosis. 

Two classification methods, U-Net and DeepLab3+, were tested for each application, with the best-performing algorithm selected for the final model. The team validated the framework using ten independent validation set images and further evaluated it on 442 melanoma WSIs from the Dutch Early-Stage Melanoma (D-ESMEL) study. 

Accuracy Across Tissue Features

The model demonstrated high performance, with median accuracy exceeding 92% for all applications. The median F1-score (Dice score) was above 80% for most applications, with only the ulceration detection application scoring lower at 70%. 

Expert dermatopathologist review of the 442 D-ESMEL images confirmed these results, with 78% of images receiving the highest rating (perfect segmentation with less than 5% error). Another 14% were considered sufficiently accurate for further analysis despite minor errors. Only 8% of images were deemed unsuitable due to either poor tissue quality (3%) or significant segmentation errors (5%). 

The researchers noted that segmentation challenges primarily occurred in cases with very thin melanomas, biopsies with minimal tumor areas, or tumors with unusually large cells.  

“By analyzing mitotic activity across the entire tumor, rather than limiting the count to visually selected hotspots, we provide a more comprehensive and objective picture of tumor proliferation,” Dr. Mooyaart noted. “This wider view may uncover patterns and regional heterogeneity in mitotic activity that are otherwise missed, potentially leading to refined stratification of tumor aggressiveness and improved staging accuracy.” 

Dr. Mooyaart also stated that the team’s approach of developing seven sequential applications, rather than a single comprehensive algorithm, enabled greater flexibility and optimization of each feature detection. This modular design also enables future refinement of specific applications without requiring retraining of the entire system, she added. 

Support for Pathologists and Enhanced Stratification

“Our model can quantify histological features such as tumor area, cell density, mitotic activity, and characteristics of the tumor microenvironment, which are often time consuming or difficult to assess manually,” explained Dr. Mooyaart. “This tool is designed to support—not replace—pathologists, allowing them to focus on higher-level diagnostic reasoning while benefiting from consistent, reproducible data extraction.” 

Dr. Mooyaart emphasized that, by stratifying patients more accurately, clinicians can tailor surveillance intensity or consider adjuvant therapies for those with aggressive tumor phenotypes, thereby optimizing outcomes and minimizing overtreatment. This personalized approach is particularly valuable for patients with thin melanomas, where current staging systems may not adequately capture the diverse biological behaviors these tumors can exhibit. 

 “Our approach has the potential to reintroduce mitotic activity as a reliable prognostic marker in thin melanoma tumors, pending validation in larger, prospective studies,” Dr. Mooyaart noted.

This is particularly significant considering that mitotic rate was removed from the AJCC staging system’s 8th edition specifically due to concerns about inter-observer variability, despite its recognized correlation with disease outcomes. 

Limitations and Future Directions

The researchers acknowledge that additional training with images obtained from a broader range of scanners would be needed to ensure the robustness and generalizability of the model across different clinical environments. 

“This was a study more showing the proof of principle,” Dr. Mooyaart said. “We would like to see which features are most predictive, which we will test in the Dutch melanoma study (D-ESMEL), and those features can then be optimized using more training data from even more scanners and different H&Es from different laboratories.” 

Additionally, the model currently struggles with distinguishing pre-existing nevi from melanoma cells, which could lead to misidentification. The researchers attempted to develop a separate application to recognize pre-existing nevi within cutaneous melanoma slides, but they were limited by the small number of available samples.  

“Detecting pre-existing nevi within melanoma slides is crucial for understanding the evolutionary trajectory from benign to malignant lesions,” noted Dr. Mooyaart. “Accurate segmentation of these components would allow us to study the histological and possibly molecular transitions associated with malignant transformation.” 

The team is now investigating which features are most predictive of outcomes in the D-ESMEL study, which includes 442 patients with primary melanoma, with 46% having stage I and 54% having stage II melanomas.  

“The quantitative features extracted by our AI model, such as mitotic activity, tumor cellularity, and stromal composition, can help identify patients who need treatment in an early stage,” said Dr. Mooyaart. “For instance, within the subset of patients with thin melanomas, our data may help identify high-risk individuals who are more likely to experience progression or metastasis.” 

This study received financial support from the KWF Dutch Cancer Society. 

References

1. Kerkour T, Hollestein L, Nigg A, et al. Automated assessment of skin histological tissue structures by artificial intelligence in cutaneous melanoma. Pathol Res Pract. 2025;269:155923. doi:10.1016/j.prp.2025.155923