AI screening detects pelvic MRI anomalies but misses hydrogel spacers

A self-supervised model detected clinical pelvic MRI anomalies with AUC 0.95, but missed all five hydrogel spacers and lacked downstream clinical validation.

KEY POINTS

  • This methodological study from Lund University and Skåne University Hospital, Sweden, developed a self-supervised deep learning framework for identifying unexpected anatomical findings and imaging artifacts before they enter automated radiotherapy workflows. The pelvic component used the LUND-PROBE dataset of 467 prostate cancer patients, divided into 384 training, 29 validation, and 54 test cases acquired using T2-weighted MRI on a single 3-T scanner. Additional evaluation included 242 synthetically generated pelvic anomalies, 36 clinical pelvic anomalies, and independently collected brain MRI datasets.
  • The framework was trained exclusively on normal images, without requiring examples of individual anomaly categories during training. A two-stage architecture combined a residual vector-quantized variational autoencoder with a vision transformer encoder and a factorized bidirectional transformer learning the distribution of normal anatomical representations. During inference, anomaly detection combined perceptual differences between observed and reconstructed or “healed” images with token-surprisal scores derived from negative log-likelihood estimates, producing both an overall anomaly score and spatial heatmaps.
  • Detection performance was high for synthetic pelvic MRI artifacts. Across 242 simulated abnormalities, including motion, noise, ghosting, Fourier-space spikes, and blurring, the model achieved AUC 0.97 (95% CI 0.96–0.99). At an operating point corresponding to a 5% false-positive rate, sensitivity was 91.0% and specificity 95.0%. However, performance differed between artifact types, with 11/49 blurred-image cases missed, compared with complete detection of the evaluated noise and ghosting cases at the selected threshold.
  • Clinical pelvic MRI anomalies were detected with AUC 0.95 (95% CI 0.90–1.00), corresponding to 90.9% sensitivity and 95.0% specificity at the same 5% false-positive operating point. The model identified all eight evaluated hip implant cases, four cervical brachytherapy applicator cases, and 14 cropped-field-of-view examinations. However, none of the five hydrogel rectal spacers was detected, demonstrating that high aggregate discrimination can conceal complete failure for a specific clinically relevant anatomical variation.
  • Performance was lower in the independent brain MRI evaluation. Using normal training data from IXI and fastMRI and 255 clinically annotated anomalous brain examinations from fastMRI+, the model achieved AUC 0.81 (95% CI 0.74–0.87). Qualitative anomaly heatmaps showed spatial correspondence with radiologist-annotated masses, resection cavities, postoperative changes, edema, and other abnormalities. However, evaluation used the axial slice containing the largest annotated abnormality rather than complete MRI examinations, potentially overestimating performance in routine unsupervised screening.
  • An illustrative MRI-only planning example demonstrated the potential consequence of undetected abnormal imaging input. In a patient with a metallic hip implant, MRI signal loss was followed by an anatomically incorrect synthetic CT reconstruction that omitted the expected high-density implant region and introduced soft-tissue-like anatomy. This example, presented to motivate the study, involved an input outside the synthetic CT software's intended use and was not part of the anomaly-detection validation. The investigation did not demonstrate whether the proposed screening model would prevent such downstream failures.
  • The study did not establish a safe operating threshold for autonomous MRI quality assurance. Pelvic MRI was acquired predominantly using a single institutional scanner and protocol, the clinical anomaly cohort was small, and detection sensitivity varied substantially between anomaly categories. The authors did not test whether flagged images actually produced inaccurate synthetic CT, segmentation, or treatment plans, nor whether unflagged images were safe for downstream processing. Prospective multicenter evaluation, scanner-specific calibration, complete-volume assessment, and quantitative analysis of downstream clinical errors remain necessary.

CLINICAL TAKEAWAY

Self-supervised anomaly detection may provide an additional quality-control layer in MRI-only and AI-assisted radiotherapy workflows by identifying unexpected anatomy or imaging artifacts before automated segmentation and synthetic CT generation. However, failure to detect certain clinical variations despite high overall AUC, together with the absence of downstream dosimetric validation, means that the framework cannot yet replace human image review or independently authorize automated planning.

SOURCE

Physics and Imaging in Radiation Oncology

Browse more research Suggest a correction