ESTRO 2024 - Abstract Book

S3046

Physics - Autosegmentation

ESTRO 2024

Accurate delineation of pelvic lymph node (LN) areas is crucial in prostate cancer radiotherapy. However, substantial variability exists in the contouring practices among radiation oncologists. This study aims to develop and validate an autosegmentation algorithm guided by established guidelines for pelvic LN areas in prostate cancer treatment.

Material/Methods:

We retrospectively acquired CT simulation images from 54 prostate cancer patients and engaged two experienced radiation oncologists to meticulously contour pelvic LN regions following the NRG International Consensus Atlas. Utilizing a 2D Unet architecture, we split the dataset into an 8:1:1 ratio for training, validation, and testing to create our autosegmentation model. Subsequently, we conducted clinical validation using a subset of CT images from an additional 98 prostate cancer patients. Radiation oncologists subjected The LN segmentation inference to critical verification and correction to ensure precision. Finally, a comprehensive quantitative assessment was conducted, encompassing Volume Dice Similarity Coefficient (VDSC), Hausdorff Distance (HD), Average Hausdorff Distance (AVHD), 95% Hausdorff Distance (95HD), Added Path Length (APL), and Surface Dice Similarity Coefficient (SDSC).

Results:

The model demonstrated a mean VDSC of 0.84 in the testing set, indicative of commendable overall performance. Across the clinical validation set, the model exhibited substantial agreement with manually corrected LN areas, with mean VDSC values ranging from 89.45% to 97.22%, as detailed in Table. Moreover, a significant correlation between VDSC and SDSC was observed ( p < 0.01). The AHD values, ranging from 3.656 mm to 15.253 mm, demonstrated reasonable spatial alignment between algorithm-generated and manually corrected contours. Remarkably, external and common iliac LN areas consistently achieved VDSC values exceeding 0.95, signifying a high level of agreement. Conversely, VDSC values exhibited greater variability for presacral and obturator lymph nodes. Notably, the autosegmentation process significantly reduced the mean time required for radiation oncologists to review and validate lymph node areas (7.3 minutes) compared to manual contouring (27.8 minutes, p < 0.01).

Conclusion:

Our findings validate the feasibility of a guideline-based autosegmentation model in aiding radiation oncologists in precisely delineating pelvic LN areas. These results hold substantial academic implications, affirming the algorithm's potential as a valuable clinical tool that can enhance treatment precision and contribute to advancements in prostate cancer management.