Patient-specific deep learning tracking for real-time 2D pancreas localisation in kV-guided radiotherapy

Abstract

Background and purpose: In pancreatic stereotactic body radiotherapy (SBRT), accurate motion management is crucial for the safe delivery of high doses per fraction. Intra-fraction tracking with magnetic resonance imaging-guidance for gated SBRT has shown potential for improved local control. Visualisation of pancreas (and surrounding organs) remains challenging in intra-fraction kilo-voltage (kV) imaging, requiring implanted fiducials. In this study, we investigate patient-specific deep-learning approaches to track the gross-tumour-volume (GTV), pancreas-head and the whole-pancreas in intra-fraction kV images.

Materials and methods: Conditional-generative-adversarial-networks were trained and tested on data from 25 patients enrolled in an ethics-approved pancreatic SBRT trial for contour prediction on intra-fraction 2D kV images. Labelled digitally-reconstructed-radiographs (DRRs) were generated from contoured planning-computed-tomography (CTs) (CT-DRRs) and cone-beam-CTs (CBCT-DRRs). A population model was trained using CT-DRRs of 19 patients. Two patient-specific model types were created for six additional patients by fine-tuning the population model using CBCT-DRRs (CBCT-models) or CT-DRRs (CT-models) acquired in exhale-breath-hold. Model predictions on unseen triggered-kV images from the corresponding six patients were evaluated against projected-contours using Dice-Similarity-Coefficient (DSC), centroid-error (CE), average Hausdorff-distance (AHD), and Hausdorff-distance at 95th-percentile (HD95).

Results: The mean ± 1SD (standard-deviation) DSCs were 0.86 ± 0.09 (CBCT-models) and 0.78 ± 0.12 (CT-models). For AHD and CE, the CBCT-model predicted contours within 2.0 mm ≥90.3 % of the time, while HD95 was within 5.0 mm ≥90.0 % of the time, and had a prediction time of 29.2 ± 3.7 ms per contour.

Conclusion: The patient-specific CBCT-models outperformed the CT-models and predicted the three contours with 90th-percentile error ≤2.0 mm, indicating the potential for clinical real-time application.

Keywords: CBCT; Deep learning; Markerless tracking; Pancreas SBRT.

Fig. 1

Illustration of workflow showing (a) population model training using CT-DRRs generated from 19 pCT scans, and (b) fine-tuning the population model into patient-specific models using CBCT-DRRs/CT-DRRs, (c) Contour prediction on the corresponding intra-fraction kV images, and (d) evaluation of prediction against the reference label.

Fig. 2

Illustration of training data creation (a) pCT/ CBCT scans were used to generate training datasets (b) labelled DRRs for three contours (DRR + contour (GTV, pancreas-head, and pancreas)), (c) Data sorting for quadrant-based training. (d) cGAN model architecture with U-net generator and patch-GAN classifier.

Fig. 3

Violin plots showing the median (solid line), interquartile range (dashed lines), and frequency width for (a) DSC, (b) absolute CE along the u-axis, (c) absolute CE along the v-axis, (d) AHD, and (e) HD95. Percentage comparisons are shown against thresholds (0.8 for DSC, 2.0 mm for CE and AHD, and 5 mm for HD95). CBCT-models are shown in blue, and CT-models in magenta. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4

Exemplar of CBCT-model prediction on kV images for (a) GTV, (b) pancreas-head, (c) pancreas, (d) when all the contours are combined. (e) and (f) show inaccurate predictions.