Geometric and dosimetric evaluation for breast and regional nodal auto-segmentation structures

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate delineation of target structures is critical for effective radiotherapy planning.
• AI-driven auto-contouring (AC) offers potential efficiency gains but requires validation against manual contouring (MC).

Purpose of the Study:

• To evaluate the geometric and dosimetric accuracy of AI-generated contours (ACs) compared to physician-drawn manual contours (MCs) for intact-breast and post-mastectomy radiotherapy plans.
• To assess the clinical applicability of AI auto-contouring in breast cancer radiotherapy.

Main Methods:

• Geometric evaluation using Dice Similarity Coefficient (DSC), mean surface distance, and Hausdorff Distance for breast, chestwall (CW), and nodal structures (axillary [AxN], supraclavicular [SC], internal mammary [IM]).
• Dosimetric evaluation by assessing dose coverage (Vx%) on ACs using MCs' plans.
• Comparison of ACs and MCs across 66 breast cancer patients for both intact-breast and post-mastectomy scenarios.

Main Results:

• Positive correlation between volume and DSC for intact-breast, AxN, and CW.
• Intact-breast ACs were dosimetrically similar to MCs, with comparable V95%.
• ACs for AxN and SC showed insignificant dosimetric differences, but IMN showed significant differences.
• Post-mastectomy ACs for AxN and SC were consistent, but IMN showed significant differences.
• 94.1% of AC-breasts met ΔV95% <5% variation when DSC > 0.7, while only 62.5% of AC-CWs achieved this.

Conclusions:

• AI auto-contouring demonstrates high accuracy for intact-breast radiotherapy planning.
• Manual adjustments may be necessary for regional nodal structures (AxN, SC) and post-mastectomy chestwall (CW) and internal mammary nodes (IMN).
• Thorough local validation of AI auto-segmentation software against institutional manual contours is essential before clinical adoption.