AI-Derived Tumor Shape Metrics Show Prognostic Potential in Melanoma

Artificial intelligence (AI) is increasingly being integrated into dermatopathology and histopathology workflows. In particular, convolutional neural networks (CNNs)—a form of machine learning designed to analyze images—have gained attention for their ability to automate time-consuming tasks and identify patterns that may not be easily recognized by the human eye. Over the past decade, these models have been applied to medical imaging across several specialties, including pathology, where the adoption of whole slide imaging has enabled digital analysis of tissue sections.¹

Despite this progress, melanoma research has lagged somewhat behind other cancers such as breast and prostate in the application of AI. One challenge is the complex and highly variable appearance of melanoma under the microscope. Another limitation has been the relatively small number of digitized melanoma specimens with long-term follow-up data needed to develop and validate predictive models.²

A recent study explored whether AI could help address this gap by analyzing melanoma tumor morphology on digital slides and identifying objective features that may predict patient outcomes. Rather than combining multiple clinical and genetic variables, the investigators focused specifically on tumor shape and size derived from automated image analysis.³

Morphology Remains Central in Melanoma Prognosis

In melanoma, morphological features of the tumor remain central to prognosis. The most widely used measure is Breslow thickness, which represents the vertical depth of tumor invasion and has been a key prognostic marker for more than 5 decades. Although highly informative, Breslow thickness is measured manually and may vary between observers.

In recent years, researchers have proposed additional morphologic features—such as width of invasion and calculated tumor area—that may improve risk stratification. However, these measurements typically require manual estimation by pathologists, which can introduce variability and increase workload.

The authors of the new study hypothesized that AI-based image analysis could provide objective measurements of tumor morphology and potentially reveal new features linked to survival.

Training a CNN to Identify Melanoma

The researchers used a custom-built convolutional neural network designed to analyze whole slide images of melanoma tissue. The model was trained to classify every pixel on a slide as either “invasive melanoma” or “not invasive melanoma,” based on manually annotated training images.

Once the tumor regions were identified, the system calculated several quantitative characteristics of the tumor’s shape and size. These included:

• Tumor area
• Tumor perimeter
• Major axis length (the longest dimension of the tumor)
• Minor axis length (the shortest dimension, representing tumor depth)
• A new measurement called the Nodularity Index (NI)

The analysis focused on the largest continuous tumor region in each specimen to ensure consistent measurements.

Introducing the Nodularity Index

One of the study’s main innovations was the Nodularity Index, a measure of tumor shape rather than size. The NI is calculated as the ratio between the tumor’s depth (minor axis length) and its width (major axis length). In practical terms, the index reflects how “rounded” or vertically oriented a tumor is.

A higher Nodularity Index indicates a tumor that grows more deeply relative to its width, which may correspond to the growth pattern seen in nodular melanoma. Importantly, this metric isolates tumor shape independently of overall size—something that is difficult to measure reliably through routine histopathologic assessment.

Study Population and Survival Analysis

The analysis included 745 melanoma cases drawn from 5 independent cohorts across the United Kingdom, Australia, and the Czech Republic. All cases had associated clinical information and survival data, with a median follow-up of 5.7 years and a maximum follow-up exceeding 30 years.

The dataset included the 4 major melanoma subtypes: superficial spreading, nodular, lentigo maligna, and acral melanoma. However, rarer subtypes were excluded.

Using statistical survival models, the investigators evaluated how each AI-derived tumor parameter related to patient outcomes.

Key Findings

Across the dataset, all 5 tumor measurements showed significant associations with survival. Larger tumors—reflected by greater area, perimeter, and axis lengths—were linked to worse overall survival and melanoma-specific survival.

The minor axis length showed a strong correlation with conventional Breslow thickness (Spearman correlation 0.89). Because of this close relationship, the researchers referred to the AI-derived measurement as digital Breslow thickness (dBT).

Digital Breslow thickness performed similarly to the traditional Breslow measurement in predicting patient outcomes. Both measures demonstrated comparable predictive accuracy for melanoma-specific survival.

The Nodularity Index also emerged as an independent predictor of survival. Tumors with higher NI values—indicating greater depth relative to width—were associated with poorer outcomes. Even after adjusting for patient age, sex, and tumor location, higher NI values remained linked to increased mortality risk.

When the Nodularity Index was added to existing prediction models, modest improvements were observed. For example, combining NI with digital Breslow thickness increased the accuracy of predicting 5-year melanoma-specific survival. Adding NI to the American Joint Committee on Cancer staging system improved predictive performance by approximately 3%.

Biological Correlations

The AI-derived measurements also aligned with known clinicopathologic features of aggressive melanoma. Tumors with larger size parameters and higher Nodularity Index values were more likely to demonstrate ulceration, vascular invasion, and higher mitotic activity. They were also more commonly associated with nodular melanoma subtype.

These findings suggest that the automated measurements capture meaningful biological characteristics of tumor growth.

Limitations and Future Directions

The study has several limitations. Although data were drawn from multiple cohorts, the final dataset contained 745 cases due to strict inclusion criteria requiring a single whole slide image per patient. The analysis was also retrospective, meaning prospective validation will be necessary before clinical adoption.

Additionally, rare melanoma subtypes such as spitzoid and desmoplastic melanoma were not included, limiting the generalizability of the findings.

Future research will focus on validating these results in larger datasets and exploring additional morphological features that may be identified through AI-based analysis.

Potential Clinical Implications

The findings suggest that AI-driven segmentation of melanoma on digital slides can generate objective measurements of tumor morphology with prognostic value. In addition to reproducing established features like Breslow thickness, the approach introduces new shape-based metrics such as the Nodularity Index.

Because these measurements are generated automatically during image analysis, they could potentially enhance prognostic assessment without adding to the pathologist’s workload. With further validation, AI-derived morphologic parameters may eventually complement traditional staging systems and provide additional insight into melanoma behavior.

References

1. Li Z, Koban KC, Schenck TL, Giunta RE, Li Q, Sun Y. Artificial intelligence in dermatology image analysis: current developments and future trends. J Clin Med. 2022;11(22):6826. Published 2022 Nov 18. doi:10.3390/jcm11226826
2. Kalidindi S. The role of artificial intelligence in the diagnosis of melanoma. Cureus. 2024;16(9):e69818. Published 2024 Sep 20. doi:10.7759/cureus.69818
3. Clarke EL, Magee D, Newton-Bishop J, et al. AI-derived prognostic biomarkers from melanoma whole slide image segmentation: an initial discovery and assessment. J Pathol Clin Res. 2026;12(2):e70075. doi:10.1002/2056-4538.70075