Abstract
Introduction: Computed tomography (CT) is one of the most important medical imaging modalities. The high patient radiation doses and frequency of examinations encountered in CT, have sparked an increased scientific and clinical interest in individual patient dose estimation. In 2014, the Council of the European Union issued Directive 2013/59/EURATOM, asking Member States to strengthen their requirements concerning patient information, including the recording and reporting of radiation doses from medical procedures. Modern computational dosimetry methods can provide accurate patient dose estimations in a non-invasive manner, using the power of modern computer systems. Also, due to the fact that they do not need to use specialized materials or consumables, computational methods have the potential to reduce the cost of individual patient dosimetry, at reasonably affordable levels. However, to date, the required know-how and computational time needed for the determination of patient dose ...
Introduction: Computed tomography (CT) is one of the most important medical imaging modalities. The high patient radiation doses and frequency of examinations encountered in CT, have sparked an increased scientific and clinical interest in individual patient dose estimation. In 2014, the Council of the European Union issued Directive 2013/59/EURATOM, asking Member States to strengthen their requirements concerning patient information, including the recording and reporting of radiation doses from medical procedures. Modern computational dosimetry methods can provide accurate patient dose estimations in a non-invasive manner, using the power of modern computer systems. Also, due to the fact that they do not need to use specialized materials or consumables, computational methods have the potential to reduce the cost of individual patient dosimetry, at reasonably affordable levels. However, to date, the required know-how and computational time needed for the determination of patient dose from CT examinations, have been discouraging researchers from pursuing this goal. Till now, the gold-standard method for individual patient CT dosimetry has been the application of Monte Carlo (MC) radiation transport methods on 3-dimensional (3D) patient specific phantoms that are created through the CT datasets of the patients themselves. The application of MC on such phantoms allows the creation of 3D dose maps. In turn, those maps can be used to calculate individual organ doses through some form of segmentation. However, as mentioned above the tools and know-how required make the wide application of such tools in clinical practice difficult. The purpose of this work was the development of an artificial intelligence (AI) assisted computational method, which will significantly contribute to making the estimation of individual patient dose in CT practically feasible at the clinic. More specifically, this work aimed to substitute the MC simulations needed in order to create the 3D dose distribution maps. Methods: A total of 95 randomly selected, anonymized CT examinations were selected from the institutional repository of the University Hospital of Heraklion. Anonymization was performed using specialized software and also manually checked to ensure it was successful, by looking into the image metadata, using a Digital Imaging and Communications in Medicine (DICOM) image viewer. Appropriate approval was obtained from the institutional review board, ensuring proper use of patient data. All CT examinations were performed on the same 64-detector row CT scanner (Revolution GSI; GE Medical Systems, Waukesha, Wisconsin, USA), using the typical adult thorax imaging protocol used in the hospital. Dose images for each CT dataset were created by using the MC-based software ImpactMC. The software runs MC simulations of radiation transport on the CT imaging datasets of individual patients and calculates the dose images. Resulting dose images correspond 1-to-1 to the CT images they were produced from, not requiring any registration. These MC-calculated dose images are used as the ground truth dose images in this study. ImpactMC requires input describing the geometric and dosimetric characteristics of the CT scanner, such as beam collimation, fan angle, spectrum of the beam and rotation time, among others. Data preprocessing was performed using Python notebooks and available, open-source, Python libraries. A total of 11424 DICOM CT images and an equal number of ground truth dose images (also in DICOM format) were perused and analyzed. To avoid fidelity loss, all images were used in their original size and no contrast enhancements (which are usually irreversible) were used. Imaging data and relevant patient information was extracted in bulk from DICOM using the Pydicom Python library and put in spreadsheets for further processing. The dataset was separated into a training dataset (77 patients, 9231 CT images and their 9231 respective ground truth dose images, created with MC methods) and a method evaluation dataset (18 patients, 2193 CT images and their 2193 respective ground truth dose images, created with MC methods). Conditional generative adversarial networks (cGANs) and more specifically image-to-image translation (pix2pix) were trained to be able to create the dose maps directly from the CT images of patients. This architecture involves two competing neural networks, namely the generator and the discriminator, working iteratively in order to learn how to create a desired type of images through looking at a different type of related images. In practice the network used in this study aimed at creating synthetic dose images by looking at how an actual pair of a CT image and its corresponding ground truth dose image (created with MC methods) looks like. The pix2pix was trained on pairs of CT and their respective, ground truth, MC-generated images of 77 randomly selected patients (47 Male, 30 Female). The total number of images used for the training was a subset of 847 from the initial 9231 images of the 77 patients that were included in the test set. Through many epochs of converging/successful training, the pix2pix model was able to learn how to create realistic dose images by just looking at CT images of patients. However, for the pix2pix cGAN to work properly, image pixel/voxel values needed to be rescaled between -1 and 1 (0 and 1 for dose images) during the data preprocessing phase, by dividing every pixel value in each image, with the highest pixel value in the image. This rescaling resulted in keeping the mathematical proportions of the images, but also to the disconnection of pixel value and the absolute dose to the tissues represented in the dose images. As a result, the synthetic dose images created by pix2pix, needed to be rescaled once more to correspond to actual dose levels. This was achieved by creating and employing a regression model that was capable of predicting the average dose value of the generated dose images. These predicted values were then used to rescale the synthetic dose images up to realistic dose levels. The regression model was trained on image data that were extracted from the DICOM headers of images, along with calculated values from the CT images themselves. DICOM header data used were reconstruction diameter, table height, X-ray tube current and slice location, while calculated quantities included patient width in the x dimension of the axial plane, patient width in the y dimension of the axial plane and water equivalent diameter (WED). The calculated variables were calculated by segmenting the CT images using computer vision (CV) methods and saving the data along with DICOM header data for every image. The evaluation of the method was two tiered. Evaluation of the dosimetric accuracy was performed by applying the trained models on images that were never seen by any of the AI models and comparing organ doses calculated on the synthetic dose images versus the corresponding doses calculated on the ground truth dose images produced with MC methods. Organ segmentation contours were applied on all slices of all 18 patients of the method evaluation dataset, and average dose to the lungs, skin, female breast, bone and heart was calculated on both the synthetic and the ground truth images, in order to be compared. The synthetic images were also evaluated for quality by using typical image comparison methods such as the structural similarity Index (SSIM), root mean squared error (RMSE), peak signal-to-noise ratio (PSNR) and spectral angle mapper (SAM). Furthermore, the gamma index pass rate (GIPR) was also used to evaluate the images in respect to their morphological and dosimetric accuracy. Results: The cGAN training for the subset of 847 images selected from the original training dataset was performed on a cloud computing platform using an instance with Graphics processing unit (GPU) acceleration (P4000 GPU). Training for 900 epochs took 52 hours to complete, while training a smaller dataset for the same number of epochs took proportionally less time, according to the smaller number of images included in the training. Nash equilibrium was achieved within about 200-400 epochs. The regression model training was performed using the PyCaret library on the dataset. Correlation heatmaps were used for the selection of the variables that would be part of the model training. The 10-fold cross-validated extra trees type model that was selected achieved a mean average error of 2.2% and did not suffer from over-fitting. The R2 value was 0.989. Using both models the creation of synthetic dose images from CT images takes seconds per image. A whole study can be processed in a few minutes and images can be saved for further processing and analysis. The dosimetric validation of the method yielded a mean absolute difference of 8.3% between doses calculated on the synthetic versus the ground truth dose images. Differences did not exceed 20% across all organs and patients. The synthetic dose images that were produced by the trained pix2pix model looked very similar to the ground truth images, however the methods applied for the quantification of their similarity yielded positive results, indicating good similarity. The average value of the more dosimetrically relevant metric of GIPR was found to be 77.7% ± 10.8%. Discussion and conclusions: This work represents a novel approach in personalized patient dosimetry in CT. The proposed method is capable of quickly calculating individual patient organ doses by using only the patient CT datasets as input and skipping the time-consuming MC dose image generation. The average absolute error of 8.3% that the method achieved is acceptable for the purposes of radiation protection in CT. Also, the method proposed herein creates synthetic dose images of acceptable quality at a fraction of the time needed for MC dose image generation. The achieved average GIPR value for the method evaluation dataset of 77.7% ± 10.8%, sets a benchmark value for synthetic image quality in CT dosimetry. By itself, the availability of good quality dose images enables further analysis for the purpose of patient dosimetry, quality and regulatory audits, or other applications such as epidemiological studies, in which organ dose calculations are pertinent. The training dataset used in this study was large enough for the AI models to be trained properly. An interesting point is that the non-uniformities and differences between the randomly selected training and evaluation datasets were not an issue for the models’ performance. Presumably, larger datasets will increase the predictive power of the models and yield even better dosimetric results, potentially covering other CT imaging protocols and X-ray energies. The time required to train the AI models was relatively long but within realistically achievable levels, making the method worth developing for individual centers, scanners and imaging protocols. The use of the pix2pix cGAN has been a good example of deep neural network architecture that is capable of figuring out an appropriate loss function without the user expressly defining it. Even if the pix2pix falls under the category of GANs and its deep learning (DL) processes are not explainable, the fact that it is a type of supervised learning makes it a little easier to understand and train using the Nash equilibrium method. The addition of the regression model for the scaling of the synthetic images using the average dose image value comprised a necessary step that connects the AI outputs to the physical reality of dosimetry. Other authors who used AI based dose prediction methods also applied similar methods to normalize their results. The proposed method yields dosimetric results that are comparable to alternative state of the art approaches, while being simple, easy to understand and elegant. Due to the use of Python programming, open-source data processing pipelines can be employed to automate and streamline steps of the process to save time, cost and effort in the clinical environment. Furthermore, the performance of the models and the overall method was not affected by the presence of different types of pathology in the images. However, further investigation of possible effects of existing pathology on the dosimetric performance would be worth exploring. Another possible future field of study would be assessing the effect of contrast agents used in the CT examinations. However, neither MC nor AI have yet produced appropriate solutions to tackle the effect of contrast agents on patient dose in CT.
show more
