Machine Learning-Based Classification of Neuroblastoma Risk Groups and MYCN Status.
In 30 neuroblastoma cases, a whole-slide-image deep learning model reportedly classified four risk/MYCN groups with 98% accuracy and MYCN amplification status with 99% accuracy, including 96% discrimination within high-risk cases.
Open original publication →What the AI sees
In 30 neuroblastoma cases, a whole-slide-image deep learning model reportedly classified four risk/MYCN groups with 98% accuracy and MYCN amplification status with 99% accuracy, including 96% discrimination within high-risk cases.
Research significance
The evidence supports a preliminary diagnostic-assistance signal rather than a treatment effect; if independently validated, histology-based AI prediction of risk group and MYCN status could inform treatment selection when molecular testing is delayed or unavailable.
Source abstract
BACKGROUND: Artificial intelligence (AI)-based methods have demonstrated considerable promise in identifying histological features in whole slide images (WSIs) and supporting risk classification in tumors. For pediatric cancers such as neuroblastoma, integrating histological findings with molecular data may enhance risk stratification and support clinical management. This study aims to evaluate the performance of a deep learning AI model on WSIs to classify neuroblastoma patients according to risk groups and determine the MYCN amplification status. METHODS: Hematoxylin and eosin-stained slides from 10 high-risk MYCN amplified, 10 high risk MYCN non-amplified, 5 intermediate-risk MYCN non-amplified, and 5 low-risk MYCN non-amplified neuroblastoma cases were digitally scanned. WSIs were processed using a deep learning-based segmentation and classification algorithms. Two patch-level datasets were constructed: 1 to predict MYCN amplification status (positive versus negative), and another to classify patients into 4 distinct risk groups (low-risk MYCN-negative, intermediate-risk MYCN-negative, high-risk MYCN-negative, and high-risk MYCN-positive). RESULTS: Our deep learning model achieved 98% accuracy of classifying neuroblastoma patients' risk groups across validation trials on slide-level classification tasks, reflecting robust performance. The accuracy of distinguishing MYCN amplified tumors from MYCN non-amplified tumors was 99%, and even within high-risk neuroblastoma cases, it was 96%. CONCLUSION: AI models can facilitate accurate risk classification of neuroblastoma patients and may also predict MYCN amplification status. They represent a promising complementary tool for pathologists working on risk stratification, especially when histological examination is limited or molecular testing for MYCN amplification is unavailable.