Uncertainty-guided AI recovers missed liver metastases during automated MRI contouring

Uncertainty-guided post-processing increased MRI-based liver metastasis detection sensitivity from 84.6% to 98.1%, while generating additional false-positive candidates.

KEY POINTS

  • This retrospective single-center study analyzed 173 patients with 224 liver metastases imaged using a 1.5-T MR-linac. The dataset was divided into training (n=100), validation (n=33) and independent testing (n=40), with 52 metastases in the test cohort.
  • The automated pipeline combined two segmentation stages: nnU-Net for the liver and treatment-relevant organs, followed by SE-ResUNet for intrahepatic lesion delineation. Uncertainty was estimated from multiple model checkpoints using predictive entropy maps, allowing cases with potentially unreliable contours or missed lesions to be flagged for additional analysis.
  • On independent testing, overall organ segmentation achieved a median Dice similarity coefficient of 0.89 and HD95 of 4.8 mm. Liver contouring performed particularly well (Dice 0.95), while stomach delineation was less accurate (Dice 0.81); lesion segmentation achieved Dice 0.86 and HD95 4.9 mm.
  • Case-level uncertainty correlated strongly with segmentation error (r = −0.802; p<0.001). The uncertainty-based triage threshold of 0.726 was selected from validation data, with an AUC of 0.96 for identifying cases containing missed metastases, supporting targeted review rather than treating every case as equally unreliable.
  • Uncertainty-guided post-processing was activated in 14 of 40 test cases (35%). It recovered seven initially missed metastases, increasing lesion detection sensitivity from 44/52 (84.6%) to 51/52 (98.1%).
  • Improved sensitivity came at the cost of 15 false-positive lesion candidates, averaging approximately 1.1 false positives per flagged case. One metastasis remained undetected because an incorrectly restricted liver contour excluded it from the lesion-search region, despite an apparently acceptable liver Dice score of 0.93.
  • Organ and lesion segmentation required median inference times of approximately 307 and 274 seconds, respectively, with additional post-processing taking 357 seconds in flagged cases. The study used one institutional MR-linac imaging protocol and did not prospectively assess contour-editing time, dose consequences, or clinical outcomes.

CLINICAL TAKEAWAY

Uncertainty-aware AI offers a clinically relevant improvement over conventional auto-segmentation by identifying suspected failures rather than simply producing contours with high average Dice scores. Recovering seven missed metastases is encouraging for MRI-guided liver planning, but false positives, a remaining critical false negative, processing time and limited external validation mean that physician review remains essential.

SOURCE

Radiotherapy and Oncology

Browse more research Suggest a correction