| name | quantum-autoencoder-mri-anomaly |
| description | Quantum autoencoder (QAE) for compression-driven anomaly detection in brain MRI. Uses angle encoding, variational encoder-decoder with trash qubits, and incompressibility-based anomaly scoring. Achieves ROC-AUC ~0.95 slice-level and ~0.813 patch-level with spatially localized anomaly heatmaps. Use when: quantum anomaly detection, brain MRI analysis, quantum autoencoder design, compression-based medical diagnostics, trash qubit encoding, variational quantum encoders. |
| metadata | {"arxiv_id":"2606.27411","published":"2026-06-25","authors":"Santanu Ganguly, Xing Liang, Dimitrios Makris","tags":["quantum-autoencoder","brain-MRI","anomaly-detection","medical-imaging","compression","trash-qubits"]} |
Quantum Autoencoder for Brain MRI Anomaly Detection
Core Concept
Anomaly detection in brain MRI using a quantum autoencoder trained to compress normal data. Anomalies (tumors, lesions) are identified by their resistance to compression relative to the learned normal manifold — higher compression resistance = higher anomaly score.
Architecture
Encoding
- Angle encoding: Maps image patches into quantum states via rotational angles
- Patch-based processing: Operates on image patches rather than full volumes
Variational Encoder-Decoder
- Encoder: Variational circuit compressing input into latent quantum state
- Trash qubits: Auxiliary qubits explicitly trained to discard information
- Decoder: Reconstructs from compressed latent representation
Anomaly Scoring
- Anomaly score = degree of input resistance to compression
- Normal data: High compression, low reconstruction error
- Anomalous data: Low compression (resists learned manifold), high reconstruction error
Key Findings
| Metric | Value |
|---|
| Slice-level ROC-AUC | ~0.95 |
| Patch-level ROC-AUC | ~0.813 |
| Baselines outperformed | Classical autoencoder, PCA |
Encoder-Decoder Asymmetry
- Effective anomaly detection arises from structured information compression within the encoder
- NOT from increased parameter magnitude or decoder expressivity
- Results in controlled compression-reconstruction trade-off with clear operating regime for principled threshold selection
Spatial Localization
- QAE produces spatially localized anomaly heatmaps aligned with tumorous regions
- Enables interpretable anomaly detection beyond binary classification
Implementation Pattern
class QuantumAnomalyDetector:
def __init__(self, n_qubits, n_trash):
self.encoder = VariationalQuantumCircuit(n_qubits)
self.decoder = VariationalQuantumCircuit(n_qubits - n_trash)
self.n_trash = n_trash
def anomaly_score(self, patch):
"""Higher score = more anomalous"""
encoded = self.encode_angle(patch)
compressed = self.encoder(encoded)
trash_state = compressed[:self.n_trash]
reconstructed = self.decoder(compressed[self.n_trash:])
return reconstruction_error(patch, reconstructed)
def encode_angle(self, patch):
angles = normalize(patch) * pi
return prepare_rotated_state(angles)
Compression-Reconstruction Trade-off
The QAE operates in a distinct regime where:
- Compression: Encoder actively discards information to trash qubits
- Reconstruction: Decoder must work with compressed representation
- Anomaly detection: Inputs outside the learned normal manifold resist compression, yielding higher reconstruction errors
This provides a principled threshold selection mechanism based on the operating regime boundary.
Applications
- Brain tumor detection and localization
- Anomaly detection in medical imaging workflows
- Decision support for radiologists
- Studying compression dynamics in quantum machine learning
Pitfalls
- Patch-level vs slice-level performance: Patch-level ROC-AUC (~0.813) lower than slice-level (~0.95) — patch granularity introduces more false positives
- DICOM dataset specificity: Evaluated on publicly available brain MRI DICOM datasets — generalization to other modalities requires retraining
- Quantum hardware constraints: Current results from simulation; actual quantum hardware may show different performance due to noise
- Threshold calibration: Anomaly score thresholds must be calibrated per dataset — no universal threshold
Activation Keywords
- quantum autoencoder, brain MRI anomaly detection, compression-driven diagnosis, trash qubit encoding, variational quantum encoder-decoder, quantum medical imaging, anomaly heatmap, incompressibility-based detection, angle encoding quantum, DICOM quantum analysis