Accelerating MRI Colon Volume Measurements and Reducing Inter-Observer Variation through Automatic Segmentation and Human-in-the-Loop Correction
The movement distribution, and volume of both chyme and gas in the colon, are important metrics to understand colonic function in health, disease, and the effects of treatments and different foodstuffs. Current methods available for assessment of these colonic contents using MRI consist mainly of manual segmentation or semi-automatic segmentation. However, these methods of segmentation are very labour intensive and too slow for clinical applications, require expert knowledge and some semi-automatic methods require use of bowel preparation. MRI scans were acquired in 2 breath holds using mDIXON sequences. We used the 'No New U-Net' (nnU-Net) ML model to automatically segment the colon, including colonic regions (ascending, transverse, descending and sigmoid-rectal). The ML-generated masks were corrected manually and the time taken for correction was recorded. ML segmentations were compared to both manual segmentations and observer corrected ML (CorrML) segmentations. Observer repeatability was also evaluated for both manual and CorrML methods to create a benchmark for the allowable error in the automatic segmentations. Analysis time was significantly reduced (p<0.0001) from 56 mins (+-11 mins (SD)) for manual masks to 11 mins (+-5 mins (SD)) for CorrML masks. Both DICE and ICC values showed excellent agreement between manual, ML and CorrML segmentations for whole colonic volume (ICC = 0.96) whilst regional volumes were good-excellent (ICC = 0.80-0.95). Inter-observer repeatability was improved when using CorrML methods over manual segmentation (ICC manual > 0.89, CorrML > 0.93). Analysis time was reduced by over 80% when using CorrML methods and whole colonic volumes measured by ML would be suitable for use with minimal checks. Hence the methods proposed here would be clinically useful.