Development and external multicentric validation of a deep learning-based clinical target volume segmentation model for whole-breast radiotherapy

Abstract

Background and purpose: In order to optimize the radiotherapy treatment and minimize toxicities, organs-at-risk (OARs) and clinical target volume (CTV) must be segmented. Deep Learning (DL) techniques show significant potential for performing this task effectively. The availability of a large single-institute data sample, combined with additional numerous multi-centric data, makes it possible to develop and validate a reliable CTV segmentation model.

Materials and methods: Planning CT data of 1822 patients were available (861 from a single center for training and 961 from 8 centers for validation). A preprocessing step, aimed at standardizing all the images, followed by a 3D-Unet capable of segmenting both right and left CTVs was implemented. The metrics used to evaluate the performance were the Dice similarity coefficient (DSC), the Hausdorff distance (HD), and its 95th percentile variant (HD_95) and the Average Surface Distance (ASD).

Results: The segmentation model achieved high performance on the validation set (DSC: 0.90; HD: 20.5 mm; HD_95: 10.0 mm; ASD: 2.1 mm; epoch 298). Furthermore, the model predicted smoother contours than the clinical ones along the cranial-caudal axis in both directions. When applied to internal and external data the same metrics demonstrated an overall agreement and model transferability for all but one (Inst 9) center.

Conclusion: . A 3D-Unet for CTV segmentation trained on a large single institute cohort consisting of planning CTs and manual segmentations was built and externally validated, reaching high performance.

Keywords: Artificial intelligence; Breast clinical target volume; Deep learning; Radiotherapy; Segmentation.

Fig. 1

Dice Similarity Coefficient (DSC), Symmetric Average Surface Distance (ASD), 95th and 100th percentile Hausdorff distance (HD_95 and HD) metrics over 300 epochs during the validation process. The corresponding bands represent the 95% confidence interval (CI) for each metric. The model converged after 298 epochs. The best metric values are listed in Supplementary Table S3. Data augmentation was not applied to the validation set shown in this plot.

Fig. 2

Nine randomly chosen patient CTs right and left samples taken at the mean cranial–caudal slices from training data of Inst 3. The clinical CTV delineation (red) is superimposed on the observed predicted (blue) segmentation ROI. The DSC and HD_95 values for each patient are shown above the images. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

DSC, HD, HD_95 and VD trend across different centers, with Inst 3 being the training center. VD negative signifies smaller CTV prediction. Box plots are shown with median (solid line) and mean (dashed lines) values. Outliers are shown with dots. Black lines show the best metric values obtained for training (applying cranial–caudal crop). Post-processing was applied to the CTV prediction as described in Materials and Methods, followed by clinical cranial–caudal cropping to enhance the main central difference.

Fig. 4

Difference in the sagittal plane of clinical and predicted contours in the caudal (left) and cranial (right) direction for all MIKAPOCo centers considered on both right (red) and left (green) breast segmented. Post-processing was applied as described in Section 2 with no extra cropping. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)