Quantum autoencoder achieves 0.95 AUC in brain MRI tumor detection

A new study from Santanu Ganguly, Xing Liang, and Dimitrios Makris explores a quantum autoencoder (QAE) for compression-driven anomaly detection in brain MRI. The approach encodes image patches into quantum states via angle encoding, then passes them through a variational encoder-decoder that intentionally discards information using auxiliary trash qubits. Normal brain patterns compress efficiently, while anomalies—like tumors—resist compression, producing higher anomaly scores. This incompressibility principle allows the model to detect deviations from learned normal manifolds without requiring labeled anomaly data.

Evaluated on publicly available brain MRI DICOM datasets, the QAE achieved a slice-level ROC-AUC of approximately 0.95 and a patch-level ROC-AUC of 0.813—outperforming classical autoencoder and PCA baselines. Analysis of learned parameters revealed a pronounced encoder-decoder asymmetry, where effective detection comes from structured compression in the encoder rather than decoder expressivity. The QAE also generates spatially localized anomaly heatmaps that align with tumorous regions, providing interpretable outputs ideal for clinical decision support. The authors highlight the controlled compression-reconstruction trade-off, enabling principled threshold selection for deployment in medical imaging workflows.