A Prior Knowledge-Guided, Deep Learning-Based Semiautomatic Segmentation for Complex Anatomy on Magnetic Resonance Imaging

Progressively refined deep joint registration segmentation (ProRSeg) of gastrointestinal organs at risk: Application to MRI and cone‐beam CT

“…For example, study by Fu et al 19 . separated stomach and duodenum into two distinct structures but combined small and large bowel into one structure, whereas we combined stomach and duodenum as one structure but separated the small and large bowels into two distinct structures, consistent with the treatment planning requirements at our institution and others 8,42 . Consistent with the findings in a prior study by Zhang et al., 12 incorporating prior knowledge from multiple preceding fractions as an ensemble segmentation improved the accuracy for all organs including the small bowel in the later treatment fractions.…”

“…Our analysis showed that ProRSeg produced reasonably accurate GI OAR segmentations exceeding a DSC of 0.80 even for challenging organs such as the small bowel. ProRSeg accuracies are better or comparable to prior published studies applied to different datasets used for MR‐Linac treatments 8,12,19 . It is notable that these published methods required ensembling of segmentations 12 from prior treatment fractions or user editing to drive semi‐automated semgentations 8 .…”

“…ProRSeg accuracies are better or comparable to prior published studies applied to different datasets used for MR‐Linac treatments 8,12,19 . It is notable that these published methods required ensembling of segmentations 12 from prior treatment fractions or user editing to drive semi‐automated semgentations 8 . ProRSeg does not require user editing or ensembling.…”

“…Highly accurate segmentation, measured as a Dice similarity coefficient (DSC) exceeding 0.8 of abdominal organs such as the liver, kidneys, spleen, as well as the stomach (excluding duodenum) has been reported by using off‐the‐shelf deep learning (DL) architectures including nnUnet 3 and Unet, 4 as well as customized dense V‐Nets 5 and new transformer based methods 6,7 applied to CT images. Slice‐wise priors provided as manual segmentations, 8 multi‐view methods using inter‐slice information from several slices and dense connections 6 as well as self‐supervised learning of transformers 9 have shown the ability to segment the more challenging GI OARs such as large and small bowel and duodenum from MRI. However, the need for manual editing 8 and large number of adjacent slices 6 required to provide priors may reduce the number of available training sets and the practicality (due to need for manual editing) of such methods.…”

Make the entrance wider and the depth deeper

“…In recent years, AI tools have emerged that require little or no operator training; these are increasingly being used in the clinic. Indeed, advances in imaging diagnostics, as well as automatic diagnosis using radiography and three-dimensional images that do not require human intervention or operation are being put to use [1]. AI may perform the functions of physicians in various fields.…”