Uncertainty & reading support

The question

How far a predicted mask can be trusted varies from slice to slice and from boundary to boundary. We show where a model hesitates as a map, and use that map to choose which slices a person should review first. We also run usability studies on how the display affects review time and missed errors.

Methods, data and evaluation

1. 20
   “Ranking slices for human review by calibrated segmentation uncertainty”

   Y. Seo, S. Oh, H. Jung, H. Do^*

   Proc. International Workshop on Imaging AI Safety, pp. 112–121, 2026.

   Int’l conferenceUncertainty & reading supportLINK full text
2. 18
   “Evaluating slice-to-slice consistency of abdominal CT organ segmentation” (in Korean)

   N. Kim, Y. Seo, H. Do^*

   Journal of the Korean Society of Medical Imaging AI, Vol. 7, No. 2, 2026.

   Domestic journalOrgan segmentationUncertainty & reading supportLINK full textPDF file
3. 16
   “Where does the model hesitate? Boundary uncertainty for organ masks”

   Y. Seo, H. Jung, H. Do^*

   Journal of Trustworthy Medical AI, Vol. 3, No. 4, 2025.

   Int’l journalUncertainty & reading supportLINK full textPDF file

People

• HDHajin DoPrincipal investigator
• YSYunjae SeoUncertainty maps for segmentation
• HJHayul JungHuman–model review workflow
• SOSerin OhUncertainty-ranked review
• SBSoeun BaeViewer usability study

Other topics

• 02Organ segmentationDelineating liver, spleen, kidneys and aorta slice by slice in abdominal CT.
• 03Low-dose reconstructionRestoring noisy low-dose slices while keeping fine structures intact.