A Pilot Study of Relating MYCN-Gene Amplification with Neuroblastoma-Patient CT Scans

Zihan Zhang, Xiang Xiang, Xuehua Peng, Jianbo ShaoarXiv 2022 · OpenReview

Abstract

Neuroblastoma is one of the most common cancers in infants, and the initial diagnosis of this disease is difficult. At present, the MYCN gene amplification (MNA) status is detected by invasive pathological examination of tumor samples. This is time-consuming and may have a hidden impact on children. To handle this problem, we adopt multiple machine learning (ML) algorithms to predict the presence or absence of MYCN gene amplification. The dataset is composed of retrospective CT images of 23 neuroblastoma patients. Different from previous work, we develop the algorithm without manually-segmented primary tumors which is time-consuming and not practical. Instead, we only need the coordinate of the center point and the number of tumor slices given by a subspecialty-trained pediatric radiologist. Specifically, CNN-based method uses pre-trained convolutional neural network, and radiomics-based method extracts radiomics features. Our results show that CNN-based method outperforms the radiomics-based method.

Introduction

Neuroblastoma is a common and often aggressive solid tumor found in infants. A key piece of the diagnostic puzzle is the MYCN gene amplification (MNA) status. If a patient is "MNA-positive," the cancer is significantly more invasive, leading to a "high-risk" classification and a poorer prognosis.

Currently, detecting MNA requires an invasive pathological examination of tumor samples (biopsy). For children, this is not only time-consuming but carries physical risks. While previous attempts have used medical imaging (Radiomics) to predict MNA non-invasively, they usually rely on "manual segmentation"—where a radiologist must painstakingly trace the 3D boundaries of a tumor. This bottleneck makes such tools impractical for fast-paced clinical environments. This study explores a way to predict MNA status using CT scans with minimal human intervention.

Methods

The study compared two primary computational paths to predict MNA status:

  1. Radiomics-based Method: Researchers extracted 107 hand-crafted mathematical features (describing intensity, shape, and texture) from the CT images. They used LASSO regression to prune these down to the most relevant features, which were then fed into standard classifiers like SVM, Random Forest, and Logistic Regression.
  2. CNN-based Method: The authors used ResNet34, a deep convolutional neural network pre-trained on ImageNet. Since ResNet expects three-channel RGB images and CT scans are single-channel grayscale, they tested three "synthesis" strategies, with stacking three identical copies of the grayscale image performing best.

The "Simplified Annotation" Innovation: Instead of full 3D segmentation, the authors only required a radiologist to provide the center point of the tumor and the number of slices it spans. The system then automatically processed the data through a specific pipeline to prepare it for the models.

Image preprocessing pipeline

The dataset consisted of retrospective CT images from 23 patients. To address the class imbalance (2 MNA-positive vs. 21 MNA-negative), data augmentation techniques such as rotation, flipping, and noise injection were employed to balance the training set.

Results and Discussion

The CNN-based approach consistently outperformed the radiomics-based models. The most significant takeaway is that precise tumor boundaries might not be necessary for genetic prediction. By using a fixed-size crop around a center point, the model still captured enough "peritumoral" context to make accurate predictions, significantly reducing the workload for radiologists.

Performance comparison of different methods

As shown in the table above, the 2nd CNN-based method (stacking identical slices) achieved the highest performance with an Accuracy of 0.79 and an AUC of 0.87. This outperformed the best radiomics-based approach (FS + Radiomics-based) which reached an AUC of 0.84. These results suggest that the low-level visual features learned from everyday objects in pre-trained models are highly transferable to medical textures, even with a limited sample size.

Conclusion

This pilot study demonstrates that deep learning can achieve high predictive accuracy for MNA status using only minimal clinician input. The proposed workflow drastically reduces evaluation time compared to traditional methods that require manual tumor tracing. While the small sample size is a limitation, the results provide a promising foundation for non-invasive risk stratification in pediatric neuroblastoma. Future work will involve validation on larger cohorts and the exploration of 3D CNN architectures.

Summary updated

Abstract

Neuroblastoma is one of the most common cancers in infants, and the initial diagnosis of this disease is difficult. At present, the MYCN gene amplification (MNA) status is detected by invasive pathological examination of tumor samples. This is time-consuming and may have a hidden impact on children. To handle this problem, we adopt multiple machine learning (ML) algorithms to predict the presence or absence of MYCN gene amplification. The dataset is composed of retrospective CT images of 23 neuroblastoma patients. Different from previous work, we develop the algorithm without manually-segmented primary tumors which is time-consuming and not practical. Instead, we only need the coordinate of the center point and the number of tumor slices given by a subspecialty-trained pediatric radiologist. Specifically, CNN-based method uses pre-trained convolutional neural network, and radiomics-based method extracts radiomics features. Our results show that CNN-based method outperforms the radiomics-based method.
BibTeX
@misc{zhang2022pilot,
  author = {Zihan Zhang and Xiang Xiang and Xuehua Peng and Jianbo Shao},
  title = {{A Pilot Study of Relating MYCN-Gene Amplification with Neuroblastoma-Patient CT Scans}},
  year = {2022},
  eprint = {2205.10619},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.10619},
}
Figure