Convolutional neural network enhancement of fast-scan low-dose cone-beam CT images for head and neck radiotherapy

Abstract

To improve image quality and CT number accuracy of fast-scan low-dose cone-beam computed tomography (CBCT) through a deep-learning convolutional neural network (CNN) methodology for head-and-neck (HN) radiotherapy. Fifty-five paired CBCT and CT images from HN patients were retrospectively analysed. Among them, 15 patients underwent adaptive replanning during treatment, thus had same-day CT/CBCT pairs. The remaining 40 patients (post-operative) had paired planning CT and 1st fraction CBCT images with minimal anatomic changes. A 2D U-Net architecture with 27-layers in 5 depths was built for the CNN. CNN training was performed using data from 40 post-operative HN patients with 2080 paired CT/CBCT slices. Validation and test datasets include 5 same-day datasets with 260 slice pairs and 10 same-day datasets with 520 slice pairs, respectively. To examine the impact of differences in training dataset selection and network performance as a function of training data size, additional networks were trained using 30, 40 and 50 datasets. Image quality of enhanced CBCT images were quantitatively compared against the CT image using mean absolute error (MAE) of Hounsfield units (HU), signal-to-noise ratio (SNR) and structural similarity (SSIM). Enhanced CBCT images reduced artifact distortion and improved soft tissue contrast. Networks trained with 40 datasets had imaging performance comparable to those trained with 50 datasets and outperformed those trained with 30 datasets. Comparison of CBCT and enhanced CBCT images demonstrated improvement in average MAE from 172.73 to 49.28 HU, SNR from 8.27 to 14.25 dB, and SSIM from 0.42 to 0.85. The image processing time is 2 s per patient using a NVIDIA GeForce GTX 1080 Ti GPU. The proposed deep-leaning methodology was fast and effective for image quality enhancement of fast-scan low-dose CBCT. This method has potential to support fast online-adaptive re-planning for HN cancer patients.

Fig. 1

U-Net architecture. Grey boxes correspond to multi-channel feature maps. The numbers of channels are shown. Arrows denote convolution layers or operations.

Fig. 2

Workflow for image pre-processing of CT and CBCT datasets.

Fig. 3

Average evaluation loss of 900–1000 epochs using three-, four- and five-group training data. X-axis denotes five models shown in the Table 1 and Y-axis denotes the loss.

Fig. 4

Quantitative measures in image quality of eCBCT and oCBCT against the corresponding CT image. (a) spider charts showing all 50 patients. Grey lines: oCBCT; Orange lines: eCBCT; (b) histogram figure showing average all 50 patients and the 10 patients with same-day images. Grey bars: oCBCT; Orange bars: eCBCT.

Fig. 5

HU Value histograms of oCBCT (grey), eCBCT (orange), and re-sim CT (blue) for all 10 testing patients (each with paired same-day CT/CBCT).

Fig. 6

The oCBCT images (left column), eCBCT images (middle column) and reference CT images (right column). (a), (b), (c) the representative image slices which have similar metrics to the average of the full testing data; (d), (e), (f) optic nerve regions; (g), (h), (i) dental regions; (j), (k), (l) parotid regions; (m), (n), (o) submandibular gland (SMG) regions; (p), (q), (r) Brainstem regions; (s), (t), (u) cord regions.

Fig. 7

HU line profiles taken at three regions at (a) Fig 6(a–c), (b) Fig 6(p–l), (c) Fig 6(s–u). The right column shows HU profiles of the red dashed lines in the left column. Y axes: HU; X axes: pixels

Fig. 8.

(a) oCBCT and processed CBCT images using (b) image-based 2D depth-5 U-Net, (c) patch-based 3D depth-5 U-Net and (d) volume-based 3D depth-3 U-Net, in comparison with (e) the reference CT.

Fig. 9.

Tumor area comparison on a testing dataset with same-day CBCT/CT: (a) oCBCT, (b) eCBCT, (c) re-sim CT, and (d) mask areas for gross tumor volumes.