Related Experiment Video
Updated: Sep 21, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
Fully Automatic Whole-Volume Tumor Segmentation in Cervical Cancer
Erlend Hodneland1,2, Satheshkumar Kaliyugarasan1,3, Kari Strøno Wagner-Larsen1,4
1Mohn Medical Imaging and Visualization Centre, Department of Radiology, Haukeland University Hospital, 5009 Bergen, Norway.
Cancers
|May 28, 2022
Summary
A new deep learning method automates uterine cervical cancer (CC) segmentation on MRI scans. While it enables automated tumor size estimation, human radiologist agreement remains superior for precise segmentation.
Area of Science:
- Oncology
- Radiology
- Medical Imaging
- Artificial Intelligence
Background:
- Uterine cervical cancer (CC) is a leading global gynecologic malignancy.
- Radiomic profiling from pelvic MRI shows promise for prognostic markers in CC treatment.
- Manual tumor segmentation is a bottleneck for clinical application of radiomics.
Purpose of the Study:
- To develop and evaluate a fully automatic deep learning (DL) method for 3D segmentation of primary CC lesions on MRI.
- To assess the performance of the DL algorithm against manual segmentations by radiologists.
Main Methods:
- A state-of-the-art deep learning algorithm was developed for 3D segmentation of primary CC tumors.
- Manual segmentations of primary tumors on T2-weighted MRI were performed by two radiologists (R1, R2) in 131 CC patients.
- The DL algorithm's performance was compared to R1/R2 using Dice Similarity Coefficients (DSCs) and Hausdorff Distances (HDs) in a test cohort (n=26).
Main Results:
- The DL algorithm achieved median DSCs of 0.60 (DL-R1) and 0.58 (DL-R2), compared to 0.78 for inter-radiologist agreement (R1-R2).
- Intraclass correlation coefficient (ICC) for tumor volume agreement was excellent between radiologists (0.93) but lower between DL and radiologists (0.43-0.44).
- The DL method successfully enabled automated estimation of tumor size and segmentation.
Conclusions:
- The developed DL algorithm facilitates automated tumor size estimation and segmentation for primary CC.
- Segmentation agreement between human radiologists is superior to the agreement between the DL algorithm and radiologists.
- Further refinement is needed to improve the DL algorithm's segmentation accuracy to match expert inter-rater reliability.
