Fig 1.
Schematic illustration of the Swin UNETR architecture used for volumetric segmentation of the prostate CTV.
Fig 2.
Training workflows for the PSN framework.
(a) PSNadaptive: the model is fine-tuned on cumulatively increasing fractions and tested on subsequent ones. (b) PSNsequence: the model is sequentially fine-tuned using the previously trained weights for each next fraction. DLg represents the pre-trained model, and n denotes the patient index (n = 1, 2,..., 5).
Table 1.
Average DSC, HD, and MSD for the Deform from ETHOS, the pre-trained Swin UNETR model, and the PSNadaptive approach using Swin UNETR, along with their standard deviations, as illustrated in Fig 2(a).
Table 2.
Average DSC, HD, and MSD for the PSNadaptive and the PSNsequence approach using Swin UNETR, along with their standard deviations, as illustrated in Fig 2.
Fig 3.
Average DSC, HD, and MSD values (± standard deviation) for prostate CTV segmentation.
Results are shown for the deformable registration method (ETHOS), pre-trained model (Swin UNETR), PSNadaptive (Fig 2(a)), and PSNsequence (Fig 2(b)).
Table 3.
Absolute deviations from physician-defined CTV dose (|ΔD95|, |ΔD98|, |ΔDmean|, |ΔD2|) across fractions.
Fig 4.
(a–b) Visual comparisons of CTV segmentation against reference contours (red), including outputs from the deformed planning CT (yellow), pre-trained model (green), and PSNadaptive (blue).
(a) Axial slice; (b) magnified view in a complex region. (c–e) 3D deviation maps between reference contours and segmentation results from the pre-trained model (c), deformed CT (d), and PSNadaptive model trained on the 1st fraction (e).