Prediction of MYCN Gene Amplification in Pediatric Neuroblastomas: Development of a Deep Learning–Based Tool for Automatic Tumor Segmentation and Comparative Analysis of Computed Tomography–Based Radiomics Features Harmonization
Ling Yun Yeow, Yu Xuan Teh, Xinyu Lu, Arvind Channarayapatna Srinivasa, Eelin Tan, Timothy Shao Ern Tan, Phua Hwee Tang, Bhanu Prakash
Abstract
OBJECTIVE: MYCN oncogene amplification is closely linked to high-grade neuroblastoma with poor prognosis. Accurate quantification is essential for risk assessment, which guides clinical decision making and disease management. This study proposes an end-to-end deep-learning framework for automatic tumor segmentation of pediatric neuroblastomas and radiomics features-based classification of MYCN gene amplification. METHODS: Data from pretreatment contrast-enhanced computed tomography scans and MYCN status from 47 cases of pediatric neuroblastomas treated at a tertiary children's hospital from 2009 to 2020 were reviewed. Automated tumor segmentation and grading pipeline includes (1) a modified U-Net for tumor segmentation; (2) extraction of radiomic textural features; (3) feature-based ComBat harmonization for removal of variabilities across scanners; (4) feature selection using 2 approaches, namely, ( a ) an ensemble approach and ( b ) stepwise forward-and-backward selection method using logistic regression classifier; and (5) radiomics features-based classification of MYCN gene amplification using machine learning classifiers. RESULTS: Median train/test Dice score for modified U-Net was 0.728/0.680. The top 3 features from the ensemble approach were neighborhood gray-tone difference matrix (NGTDM) busyness, NGTDM strength, and gray-level run-length matrix (GLRLM) low gray-level run emphasis, whereas those from the stepwise approach were GLRLM low gray-level run emphasis, GLRLM high gray-level run emphasis, and NGTDM coarseness. The top-performing tumor classification algorithm achieved a weighted F1 score of 97%, an area under the receiver operating characteristic curve of 96.9%, an accuracy of 96.97%, and a negative predictive value of 100%. Harmonization-based tumor classification improved the accuracy by 2% to 3% for all classifiers. CONCLUSION: The proposed end-to-end framework achieved high accuracy for MYCN gene amplification status classification.