Related Experiment Video
Updated: May 25, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Evaluation of segmentation accuracy and the improvement of time effectiveness using deep learning-based segmentation
Tetsu Nakaichi1,2, Yuhei Shimizu3, Satoshi Nakamura4,5,6
1Section of Radiation Safety and Quality Assurance, National Cancer Center Hospital, 5-1-1 Tsukiji, Chuo-ku, Tokyo, 104-0045, Japan.
Background:
The efficacy of deep learning-based artificial intelligence segmentation (AI-seg) in 177Lu-DOTATATE dosimetry remains underexplored. This study evaluates AI-seg's contouring accuracy, dosimetric reliability, and time efficiency.
Methods:
We analyzed 23 patients treated with 177Lu-DOTATATE. Four medical physicists (MPs) and four radiological technologists (RTs) manually delineated liver (including lesions), spleen, and kidneys on CT images. The most experienced MP modified AI-seg outputs to establish a reference contour, confirmed by a board-certified physician. Both the manual and AI-seg contours were checked against this reference using the Dice similarity coefficient (DSC), Hausdorff distance (HD), and mean distance to agreement (MDA). Maximum, mean, and minimum absorbed doses were calculated from a single time-point image using SurePlan™ MRT (MIM Software Inc.) to assess dosimetric implications. The total time required for dosimetry-including image reconstruction, segmentation, and absorbed dose calculation-was compared between the manual and reference methods.
Results:
Only one liver delineation by an MP required correction. Median (IQR) DSCs for liver, spleen, and kidneys were 0.955 (0.950-0.961), 0.929 (0.898-0.944), and 0.929 (0.918-0.942) for manual, 0.988 (0.946-0.998), 0.965 (0.858-0.995), and 0.994 (0.985-0.999) for AI-seg, respectively. All organs met the acceptable DSC threshold ([Formula: see text]0.8). Median HD in liver exceeded 10.0 mm for both methods, with AI-seg exceeding 30.0 mm in a few cases (liver: 3/23; spleen: 2/23) due to hepatomegaly. The median MDA in manual and AI-seg was below 2.0 mm. Median mean absorbed doses in the reference was 4.03 (1.69-5.81) Gy for liver, 1.55 (1.30-2.50) Gy for spleen, and 2.03 (1.45-2.44) Gy for kidneys. Only the kidneys absorbed dose from AI-seg differed significantly, showing a 1.4% increase. Dosimetry using the reference method took significantly less time than the manual approach (47.0 [30.0-58.0] min vs. 54.3 [49.5-67.0] min, p = 0.014).
Conclusions:
AI-seg with minor adjustment enables faster yet accurate absorbed dose estimation in 177Lu-DOTATATE. Despite segmentation challenges in hepatomegaly cases, MPs and RTs demonstrated competent contouring performance, supporting collaborative dosimetry workflows with physician oversight.
