Automated segmentation in pelvic radiotherapy: A comprehensive evaluation of ATLAS-, machine learning-, and deep...

Area of Science:

• Medical imaging analysis
• Radiotherapy workflow optimization
• Artificial intelligence in healthcare

Background:

• Manual segmentation of pelvic structures in radiotherapy is time-consuming and prone to variability.
• Automated segmentation using artificial intelligence (AI) offers potential for standardization and efficiency.

Purpose of the Study:

• To evaluate the performance of four automated segmentation tools on computed tomography (CT) images for pelvic radiotherapy.
• To compare atlas-based and machine learning algorithms, including deep learning (DL), for segmenting cervical and prostate cancer cases.

Main Methods:

• Retrospective analysis of 40 cervical and 40 prostate cancer CT datasets.
• Comparison of STAPLE, Random Forest, MVision (DL), and LimbusAI (DL) against manual segmentation (Ground Truth).
• Evaluation using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Distance-to-Agreement Portion (DAP), and time efficiency.

Main Results:

• Deep learning tools significantly outperformed other methods in cervical cancer CTs, showing superior quantitative metrics, qualitative scores, and reduced correction times.
• For prostate cancer CTs, all analyzed tools demonstrated comparable quantitative and qualitative performance.
• Anatomical variability in cervical cancer and standardized patient preparation in prostate SBRT may explain performance differences.

Conclusions:

• AI-driven automated segmentation, particularly DL, can enhance efficiency and accuracy in pelvic radiotherapy planning, especially for complex cases like cervical cancer.
• Further development and validation of AI tools are crucial for widespread adoption in routine radiotherapy workflows.