Generation of deep learning based virtual non-contrast CT using dual-layer dual-energy CT and its application to planning CT for radiotherapy

Abstract

This paper presents a novel approach for generating virtual non-contrast planning computed tomography (VNC-pCT) images from contrast-enhanced planning CT (CE-pCT) scans using a deep learning model. Unlike previous studies, which often lacked sufficient data pairs of contrast-enhanced and non-contrast CT images, we trained our model on dual-energy CT (DECT) images, using virtual non-contrast CT (VNC CT) images as outputs instead of true non-contrast CT images. We used a deterministic method to convert CE-pCT images into pseudo DECT images for model application. Model training and evaluation were conducted on 45 patients. The performance of our model, 'VNC-Net', was evaluated using various metrics, demonstrating high scores for quantitative performance. Moreover, our model accurately replicated target VNC CT images, showing close correspondence in CT numbers. The versatility of our model was further demonstrated by applying it to pseudo VNC DECT generation, followed by conversion to VNC-pCT. CE-pCT images of ten liver cancer patients and ten left-sided breast cancer patients were used. A quantitative comparison with true non-contrast planning CT (TNC-pCT) images validated the accuracy of the generated VNC-pCT images. Furthermore, dose calculations on CE-pCT and VNC-pCT images from patients undergoing volumetric modulated arc therapy for liver and breast cancer treatment showed the clinical relevance of our approach. Despite the model's overall good performance, limitations remained, particularly in maintaining CT numbers of bone and soft tissue less influenced by contrast agent. Future research should address these challenges to further improve the model's accuracy and applicability in radiotherapy planning. Overall, our study highlights the potential of deep learning models to improve imaging protocols and accuracy in radiotherapy planning.

Fig 1. CIRS phantom CT images.

The CT images of the CIRS phantom (Model 062M, Sun Nuclear Corp. FL, USA) acquired by the DECT (iQon Spectral CT, Philips, the Netherlands) (A) and planning CT scanner (Brilliance Big Bore (BBB) CT, Philips, the Netherlands) (B). CT, computed tomography; DECT, dual-energy computed tomography.

Fig 2. VNC-Net architecture.

The model takes contrast enhanced dual-energy CT (CE DECT) images as input and generates virtual non-contrast (VNC) images as output. The numbers denoted below the encoder/decoder blocks represent the number of channels. The bottom of the figure depicts the structure of the down-sampling block of the encoder, which comprises a UnetConv2D block and a max pooling layer. The UnetConv2D component includes two ConvBNReLU blocks, each consisting of a Conv layer that executes 2D convolutional operations, followed by batch normalization and a ReLU activation function.

Fig 3. Proposed method’s pipeline.

This is the tree diagram of the proposed method’s pipeline.

Fig 4. VNC-Net model results for DECT dataset.

Three samples from the test set are visualized. The rightmost column represents the difference map between the target and model output, and it is overlaid on the model output in the third column to highlight where the anatomy differs from the target. The images in the middle row have the right atrium (marked by the upper left yellow box of each image), the descending aorta (marked by the lower right yellow box of each image), and the rib (marked by the green box of each image) as the ROIs. DECT, dual-energy computed tomography; ROI, region of interest.

Fig 5. Linear relationship between the CT numbers from pCT and DECT.

The linear relationship obtained from the pixel values of the pCT (Brilliance Big Bore (BBB) CT, Philips, the Netherlands) and DECT (iQon Spectral CT, Philips, the Netherlands) images of the CIRS phantom (Model 062M, Sun Nuclear Corp., FL, USA) with all tissue plugs. We obtained two linear equations for values below or equal to 0 HU and above 0 HU. Magenta and blue represent pixel values from the same tissue plug at different locations, while black represents the mean value plotted, with error bars indicating the standard deviation. The error bars are not clearly discerned from the data points due to their small standard deviations. DECT, dual-energy computed tomography; HU, Hounsfield unit; pCT, planning computed tomography.

Fig 6. Conversion between pseudo DECT and pCT.

CE-pCT (A), pseudo DECT (B), pseudo VNC DECT (C), and VNC-pCT (D). To ensure the anatomy is clearly visible, CE-pCT and VNC-pCT were enlarged to match the body size of pseudo DECT and pseudo VNC DECT. CE-pCT, contrast enhanced planning computed tomography; DECT, dual-energy computed tomography; pCT, planning CT; VNC DECT, virtual non-contrast dual-energy CT; VNC-pCT, virtual non-contrast planning computed tomography.

Fig 7. ROI analysis of the VNC-pCT results.

In the top row, the ROI on the left side of each image is the liver, and the ROI on the right side is the kidney. In the middle row, the ROI in the upper left side of each image is the right atrium, and the ROI in the lower right side is the descending aorta. Lastly, in the bottom row, the ROI marked with a green box represents the ribs, while among the two ROIs marked with yellow boxes, the one on the left is the lung and the one on the right is the breast. ROI, region of interest; VNC-pCT, virtual non-contrast planning computed tomography.