Can Deep Learning-Based Auto-Contouring Software Achieve Accurate Pelvic Volume Delineation in Volumetric...

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Radiotherapy is a primary treatment for prostate cancer (PC).
• Accurate delineation of organs at risk (OARs) is vital for optimizing treatment efficacy and minimizing side effects.
• Manual segmentation of OARs is labor-intensive and subject to inter-observer variability.

Purpose of the Study:

• To evaluate the performance of Limbus Contour (LC), a deep learning auto-contouring software, for delineating pelvic structures in PC patients.
• To compare the accuracy of LC's auto-contoured OARs against manual delineations by radiation oncologists.

Main Methods:

• LC software was used to auto-contour pelvic OARs, including the bowel bag, bladder, rectum, sigmoid colon, and pelvic lymph nodes.
• The performance was assessed in 52 PC patients.
• Quantitative metrics were used to compare auto-contoured structures with manually delineated ones.

Main Results:

• LC demonstrated high accuracy for bladder (median Dice: 0.95) and rectum (median Dice: 0.83) delineation.
• Performance was limited for the bowel bag (median Dice: 0.64) and sigmoid colon (median Dice: 0.6), with instances of irrelevant structure inclusion.
• Pelvic lymph node delineation showed acceptable median Dice (0.73) but lacked sub-regional differentiation.

Conclusions:

• LC shows potential for automating OAR delineation in prostate radiotherapy, especially for the bladder and rectum.
• Enhancements are required for bowel bag, sigmoid colon, and lymph node auto-contouring, particularly regarding sub-regional accuracy.
• Further validation across larger and more diverse patient cohorts is recommended to confirm generalizability.