Session Item

Artificial intelligence (AI) has been applied to several aspects of radiation oncology practice but its role in brachytherapy is scarcely discussed. On the other hand, there has been an interest renewal in brachytherapy as a crucial part of curative radiation therapy for cervix cancer. Manual contouring in gynecological intrauterine brachytherapy (IB) treatment planning is critical and time-consuming due to a new implant for each fraction, generally once a week for 4-5 weeks. This study evaluated the feasibility of atlas-based auto-segmentation (ABAS, Elekta, Stockholm) in the delineation of rectum and bladder in cervical cancer patients during IB planning.

At our Institution, the IB procedure requires empty rectum and bladder filling (100 ml of saline solution) through a urethral catheter before each CT scan and treatment fraction. Ten different IB patients treated in 2020 have been retrospectively selected. Contours for bladder and rectum, carefully defined by two skilled radiation oncologists, were used to create an atlas library in ABAS. Organs at risks (OARs) contours were generated in ABAS on 16 planning CTs of 4 IB treated patients (January-April 2021). The contour comparison was performed using the Hausdorff distance (HD), the Mean Distance to Agreement (MDA), the Dice similarity Index (DI), and the Jaccard Index (JI). The results were compared performing the Wilcoxon Mann Whitney test to assess statistical significance (α=0.05).

The auto-segmentation process requested about 4 minutes for each CT set. Figures 1 and 2 show the geometrical metric results. ABAS best results have been recorded in the bladder contouring. The median HD, MDA, DI, and JI values for the rectum were 24.27 mm [13.04 - 35.20 mm], 4.55 mm [1.96 - 8.59 mm], 0.61 [0.30 - 0.79], and 0.44 [0.18 - 0.65], respectively; while the median metric values for the bladder were 12.18 mm [6.00 - 27.06 mm], 1.53 mm [0.87 - 2.67 mm], 0.89 [0.79 - 0.94], and 0.81 [0.65 - 0.89], respectively. The bladder and rectum results showed a statistically significant difference (p << 0.05) for all the studied metrics. The mean standard deviation of HD, MDA, DI, and JI metrics for a single patient was 5.23 mm, 1.08 mm, 0.07, and 0.07 for the rectum, while the values for the bladder were 3.43 mm, 0.31mm, 0.02, and 0.03, respectively.

In these preliminary results, bladder segmentation showed a higher accuracy compared to rectum segmentation with smaller inter-fractions variation for a single patient. The suboptimal results for the rectum could partially be explained by different organ preparation: the emptying of the rectum was self-regulated by patients at home, while the bladder was filled by a standard procedure in our Department. On the other hand, intra-patient variability is limited. The implementation of ABAS with manual adjustments in IB daily clinical practice could reduce time consumption, however, further implementations and investigations of AI are needed.