AI Image Analysis–A New Tool in CNS Tumour Diagnostics

This article was originally published in The Evidence Base for Scientific Writers Ltd

New deep learning model predicts CNS tumour DNA-methylation profile

DNA methylation profiling is used to diagnose and classify central nervous system tumours; however, this method is both labour-intensive and time-consuming, presenting challenges in a field where rapid decision-making is crucial.

To improve wait times for tumour diagnostics, researchers have developed a deep learning model called DEPLOY (Deep lEarning from histoPathoLOgy and methYlation) which can scan histopathology slides of tumour biopsy tissue samples and sort them into 10 distinct categories.

A new approach in tumour diagnostics

DEPLOY represents a significant advancement in CNS tumour diagnostics by integrating deep learning with existing clinical practices. The DEPLOY model was developed using three models: a direct model which predicts the tumour type based on histopathology images, an indirect model which guesses DNA methylation patterns based on those same images, and a demographic model which considers patient age, gender, and where in the body the biopsy was taken. The combination of these three models contribute to DEPLOY’s final classification label.

The model was trained using images from samples with known DNA methylation and tumour types. The training dataset consisted of samples from 1796 patients and was further validated on three additional datasets consisting of 360 patients. The DEPLOY model achieved a balanced accuracy of 91%.

Streamlining clinical workflows

One of DEPLOY's greatest strengths is its ability to integrate into existing clinical workflows. Since it relies on histopathology—a standard diagnostic tool in most hospitals—DEPLOY can provide rapid and reliable results without the need for additional infrastructure. This ability makes DEPLOY particularly valuable in resource-limited settings such as rural or low-income areas, where advanced DNA sequencing technologies may not be accessible.

DNA methylation profiling and histopathology analysis require skilled technicians and clinicians and can take several weeks to complete. The DEPLOY model can diagnose tumour types based on histopathology images alone—cutting down on both the time it takes to get a diagnosis and the person-power needed for each diagnosis.

Not so fast

Despite promising results, DEPLOY does have limitations. The model's accuracy is limited by the quality of the original histopathology images and the availability of demographic data. Additionally, while the model performs well in distinguishing between major tumour types, the effectiveness in classifying rare or atypical tumours may be limited. The authors of this paper did not provide the relative ratios for each tumour type; validation is still needed for other tumour types.

Why does it matter?

CNS tumours are some of the fastest growing and most aggressive, with the shortest survival time. Fast and accurate diagnosis of these tumours is critical, with early diagnosis providing patients with better outcomes. Accurate AI diagnostic tools also mean that under-resourced areas don’t have to rely on skilled lab technicians to complete the work.

Take home messages

1. DEPLOY was developed to classify CNS tumours based on histopathology images

2. The DEPLOY model is faster and more cost-effective than other methods

3. Further validation is needed—especially in rare tumour types

---

This article was written as part of a series of 'journal club' summaries for Scientific Writers Ltd and is based on the following publication.

Title: Prediction of DNA methylation-based tumor types from histopathology in central nervous system tumors with deep learning.

First Author: Hoang, D., et al

Journal: Nature Medicine

Date online: 17 May 2024

Other references:

Fagan, J. AI tool speeds up brain tumor classification. Neuroscience News. May 17, 2024. Accessed October 2024.