Evaluation of deep learning-based target auto-segmentation for Magnetic Resonance Imaging-guided cervix brachytherapy

Abstract

Background and purpose: The target structures for cervix brachytherapy are segmented by radiation oncologists using imaging and clinical information. At the first fraction, this is performed manually from scratch. For subsequent fractions the first fraction segmentations are rigidly propagated and edited manually. This process is time-consuming while patients wait immobilized. In this work, we evaluate the potential clinical impact of using population-based and patient-specific auto-segmentations as a starting point for target segmentation of the second fraction.

Materials and method: For twenty-eight patients with locally advanced cervical cancer, treated with MRI-guided brachytherapy, auto-segmentations were retrospectively generated for the second fraction image using two approaches: 1) population-based model, 2) patient-specific models fine-tuned on first fraction information. A radiation oncologist manually edited the auto-segmentations to assess model-induced bias. Pairwise geometric and dosimetric comparisons were performed for the automatic, edited and clinical structures. The time spent editing the auto-segmentations was compared to the current clinical workflow.

Results: The edited structures were more similar to the automatic than to the clinical structures. The geometric and dosimetric differences between the edited and the clinical structures were comparable to the inter-observer variability investigated in literature. Editing the auto-segmentations was faster than the manual segmentation performed during our clinical workflow. Patient-specific auto-segmentations required less edits than population-based structures.

Conclusions: Auto-segmentation introduces a bias in the manual delineations but this bias is clinically irrelevant. Auto-segmentation, particularly patient-specific fine-tuning, is a time-saving tool that can improve treatment logistics and therefore reduce patient burden during the second fraction of cervix brachytherapy.

Keywords: Automatic target segmentation; Brachytherapy; Cervical cancer; Patient-specific fine-tuning.

Fig. 1

Pairwise geometric comparisons between the auto and the respective edited segmentations and between the edited and the clinical structures. The mean CI resulting from the IOV analysis performed in is added for reference. The values of the tolerance parameter τ used in the calculation of the surface Dice are indicated under the structure names. *p < 0.05.

Fig. 2

Examples of: a) GTV-T_res, b) CTV-T_HR and c) CTV-T_IR segmentations on the sagittal T2-weighted images of three different patients (cropped views). The population-based and the patient-specific structures are shown on the top and bottom rows, respectively. The auto and the edited segmentations are shown in light and dark colors, respectively. The corresponding clinical structures are shown in orange for reference.

Fig. 3

Dosimetric comparison between the edited and the clinical structures for the two auto-segmentation methods. The [−2σ_IOV, +2σ_IOV] range, determined from the data reported in , is shown as reference for the $\mathrm{\Delta }{\mathrm{D}}_{90}^{\mathrm{r}\mathrm{e}\mathrm{l}}$ and $\mathrm{\Delta }{\mathrm{D}}_{100}^{\mathrm{r}\mathrm{e}\mathrm{l}}$. For the $\mathrm{\Delta }{\mathrm{D}}_{98}^{\mathrm{r}\mathrm{e}\mathrm{l}}$ there was no IOV data available. *p < 0.05.

Fig. 4

Time spent manually segmenting the three target structures when using the auto (population-based) and the auto (patient-specific) segmentations as a starting point and during the current clinical workflow. *p < 0.05.

Fig. 5

Total APL between the auto and the corresponding edited segmentations and the editing time. The Pearson correlation coefficients (R) are color-coded according to the scatterplot legend. For the determined R values, the corresponding p-values were smaller than 10⁻³.