3D Reasoning for Unsupervised Anomaly Detection in Pediatric WbMRI

Alex Chang, Vinith M. Suriyakumar, Abhishek Moturu, James Tu, Nipaporn Tewattanarat, Sayali Joshi, Andrea Doria, Anna GoldenbergarXiv 2021 · arXiv

Introduction

Detecting cancerous lesions in pediatric whole-body MRI (wbMRI) presents a significant "needle in a haystack" challenge. The rarity of pediatric cancer syndromes leads to a severe lack of labeled data, precluding the use of standard supervised deep learning. Furthermore, pediatric imaging is complicated by continuous anatomical growth—negating the use of standard atlases—and technical noise. Most critically, traditional unsupervised models treat MRI slices as independent 2D images, ignoring the inherent 3D structure of biological tissues and lesions.

Methodology

To address these challenges, the authors propose an unsupervised approach using a Conditional $\beta$-Variational Autoencoder ($\beta$-VAE). The model is trained exclusively on healthy scans to learn a representation of normal anatomy; anomalies are then identified by the model's failure to reconstruct them accurately.

Architecture of the Conditional $\beta$-VAE model using multi-channel sliding windows and spatial conditioning.

The framework incorporates several technical innovations:

  • Multi-channel Sliding Window: The model processes $N$ consecutive slices simultaneously, enabling "3D reasoning" without the complexity of full 3D convolutions.
  • Spatial and Metadata Conditioning: To account for anatomical heterogeneity, the VAE is conditioned on window coordinates ($w_y, w_z$) and patient-specific features such as age, sex, and weight.
  • Intensity-based Post-processing: On STIR (Short Tau Inversion Recovery) sequences, lesions typically appear hyperintense. The detection logic specifically flags areas where the reconstruction is significantly darker than the original input.

The study utilized 535 healthy wbMRI scans for training and evaluated performance on 27 real cases with annotated lesions, supplemented by synthetic tumor data.

Results and Discussion

The transition from 2D to 3D context proved transformative. Increasing the input from a single slice to a 9-slice window with spatial conditioning improved the Area Under the Precision-Recall Curve (AUPRC) from 0.033 to 0.179.

Qualitative comparison of anomaly detection performance across different slice counts and $\beta$ values.

Key Insights

  • Spatial Context: Multi-slice inputs allow the model to distinguish between normal structures (like vessels) and suspicious nodules by checking consistency across the Z-axis. As shown in the qualitative results, increasing slice count leads to significantly cleaner anomaly masks.
  • The Metadata Paradox: While sex-based conditioning improved performance, age and weight introduced noise, likely due to the dataset's limited size relative to the complexity of continuous growth patterns.
  • Hyperparameter Tuning: The $\beta$ parameter is critical; it balances reconstruction fidelity against latent space regularity. An optimal $\beta$ prevents the model from "memorizing" anomalies while maintaining sharp reconstructions.

The optimized configuration—using 5 slices, sex conditioning, and tuned $\beta$—achieved an AUPRC of 0.235 and an AUROC of 0.941. Visualizations confirm that 3D-aware models effectively suppress false positives from normal anatomical variations that often trip up 2D baselines.

Conclusion

This work demonstrates that 3D spatial context and spatial conditioning are essential for effective unsupervised anomaly detection in complex pediatric datasets. Future research will focus on scaling to larger multi-center datasets and improving the integration of continuous developmental features to better handle the spectrum of pediatric growth.

Summary updated

Abstract

Modern deep unsupervised learning methods have shown great promise for detecting diseases across a variety of medical imaging modalities. While previous generative modeling approaches successfully perform anomaly detection by learning the distribution of healthy 2D image slices, they process such slices independently and ignore the fact that they are correlated, all being sampled from a 3D volume. We show that incorporating the 3D context and processing whole-body MRI volumes is beneficial to distinguishing anomalies from their benign counterparts. In our work, we introduce a multi-channel sliding window generative model to perform lesion detection in whole-body MRI (wbMRI). Our experiments demonstrate that our proposed method significantly outperforms processing individual images in isolation and our ablations clearly show the importance of 3D reasoning. Moreover, our work also shows that it is beneficial to include additional patient-specific features to further improve anomaly detection in pediatric scans.
BibTeX
@misc{chang2021d,
  author = {Alex Chang and Vinith M. Suriyakumar and Abhishek Moturu and James Tu and Nipaporn Tewattanarat and Sayali Joshi and Andrea Doria and Anna Goldenberg},
  title = {{3D Reasoning for Unsupervised Anomaly Detection in Pediatric WbMRI}},
  year = {2021},
  eprint = {2103.13497},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.13497},
}
Figure