MyT3 9 - Artificial Intelligence and Machine Learning

Purpose:

Bone marrow is the primary site of skeletal metastases in prostate cancer. ¹⁸F-sodium fluoride (NaF) can detect activity due to malignancy, but also identifies irrelevant degenerative cortical uptake. Normal radiotracer activity in solely the marrow has yet to be described and could be the first step towards automated tumour burden calculation. We aimed to investigate the normal activity of ¹⁸F-NaF in whole bone and bone marrow in patients with localised prostate cancer.

Methods and materials:

¹⁸F-NaF PET/CT scans from 87 patients with high-risk prostate cancer from two centres were retrospectively analysed. All patients had a recent negative or inconclusive bone scan. In the first centre, a PET scan was acquired 1-1.5 hours after i.v. injection of 4 MBq/kg ¹⁸F-NaF on an integrated PET/CT system (Gemini TF, Philips Medical Systems) (53/87). In the second centre, scanning was performed 1 hour after i.v. injection of 3 MBq/kg ¹⁸F-NaF on an integrated PET/CT system (Discovery ST, GE Healthcare) (34/87). CT scans were obtained in immediate connection to the PET scan. Automated segmentation of vertebrae, pelvis, femora, humeri and sternum were performed in the CT scans using a deep learning-based method. Bone <7 mm from skeletal surfaces was removed to isolate the marrow. SUV was measured within the remaining area in the PET scan.

Results:

SUV_max and SUV_mean in the whole bone and bone marrow of the different regions were presented.

Conclusion:

We present a deep-learning approach for evaluation of normal radiotracer activity. Knowledge about radiotracer uptake in normal bone prior to cancerous involvement is a necessary first step for subsequent tumour assessment and could be of value when implementing future tracers.

Limitations:

n/a

Ethics committee approval

Lund, Sweden (552/2007).

Funding:

This study was funded by Knut and Alice Wallenberg Foundation.

Purpose:

Legislation regarding radiation dose registration creates the need for more accurate dose information. In interventional radiology, the patient dose and especially skin dose is of particular interest. Current calculation methods are based solely on the radiation dose structured report (RDSR) data or indirect dose metrics resulting in high inaccuracies. In this work, a more accurate and fully automated system is developed allowing for accurate calculations giving almost real-time feedback to the physician.

Methods and materials:

An automated dose calculation system based on Monte Carlo simulations and computer vision were developed using three RGB-D cameras (Microsoft Kinect). The calibrated camera system determined the geometric relationship between the patient and x-ray source and combines this information with the RDSR data to create a fast Monte Carlo model and calculate the patient’s skin dose. The patient is modelled using a family of mathematical (XCAT) phantoms. The system was tested on a clinical machine using an anthropomorphic RANDO-phantom and GafChromic film.

Results:

Fully automated skin dose calculations can be performed under 10 seconds within a 10% uncertainty for all field sizes, and uncertainties below 5% for field sizes smaller than 15 cm x 15 cm. Initial tests show differences between film and simulations smaller than 5%. Organ doses can be calculated offline.

Conclusion:

The built prototype shows that the combination of Monte Carlo simulation and computer vision has the potential to improve the accuracy of dose calculations in radiology. Even providing near the real-time skin dose feedback during interventional procedures allowing for the prevention of radiation-induced skin lesions. The system can also calculate organ doses to estimate the long term radiation effect.

Limitations:

Only phantom measurements were performed.

Ethics committee approval

n/a

Funding:

This work was partially funded by an EU Interreg Crossroads2 grant.

Purpose:

The aim of this study was to automatically assess the body weight prior to CT examinations.

Methods and materials:

Body weight of 100 consecutive adult patients with an indication for CT was visually estimated by a technologist and a radiologist prior to the examination. The patients were additionally asked for their current anamnestic values before the true values were weighed. A roof-mounted 3D camera was used after patient positioning on the scanner's table to fit in an individual avatar using machine learning algorithm that was taught by the 200 preceding patients. Body weight was mathematically derived from this information.

Results:

Mean body weight of our collective was 80 kg (± 19kg). The vast majority was examined wearing street-wear (94%), only few patients were wearing medical gowns (6%). Shoes were worn in 60% of the cases. Anamnestic values had a very low error from the ground truth (± 2%). Automated calculations (± 3%) were comparably precise, but estimations by radiologists (± 8%) and technologist (± 9%) were significantly worse. All assessment techniques tended to slightly underestimate the true values.

Conclusion:

Automated weight assessment is feasible in a clinical routine setting with a high precision using a 3D camera and machine learning algorithms. This information could be used for automated contrast agent adaptations in future CT generations.

Limitations:

Different interfering circumstances like large clothes, covers, positioning and other medical devices may bias the assessment of body weight by a 3D camera. The accuracy of the algorithm could be improved especially in these patients by increasing the number of cases in future projects.

Ethics committee approval

The study was approved by the local review board, written informed consent was obtained from each patient.

Funding:

The study was funded by Siemens Healthcare GmbH.

Purpose:

We aimed to assess the utility of an automatic post-processing and reporting system based on CAD-RADS^TM in suspected coronary artery disease patients.

Methods and materials:

The model was designed for CAD-RADS assessment categories with automatic coronary tree segmentation and stenosis detection algorithm based on convolutional neural networks with the training of consecutive 2000 CCTA examinations. The diagnostic value of CAD-RADS classification, one-vessel CAD, two-vessel CAD, left main CAD and three-vessel CAD were performed by the model compared to radiologists with commercially-available automated segmentation and manual post-processing and also compared to invasive coronary angiography (ICA).

Results:

Of the 322 patients in the study, 15, 9, 48, 96, 99, 35, 20 were classified as CAD-RADS 0, 1, 2, 3, 4A, 4B and 5 based on the model. The consistency test showed that the Kappa value of the model and radiologists was 0.639 (P<0.05). The Kappa value for diagnosing one-vessel CAD, two-vessel CAD, left main CAD and three-vessel CAD between the model and radiologists with CCTA is 0.758, 0.747, 0.869 and 0.596, respectively. The Kappa value for diagnosing one-vessel CAD, two-vessel CAD, left main CAD and three-vessel CAD between the model and ICA is 0.581, 0.494, 0.758 and 350, respectively. However, there is a poor agreement for detecting plaque characteristics between the model and radiologists with CCTA (Kappa value 0.349).

Conclusion:

The CNN-based CAD-RADS in CCTA images is good consistency with the radiologists. However, the poor agreement for detecting plaque characteristics is remarkable in the study.

Limitations:

This study is limited by the select bias and the fact that it is a retrospective study.

Ethics committee approval

The present study was approved by the institutional review board of the Central Hospital of Wuhan.

Funding:

No funding was received for this work.

Purpose:

The purpose of the study is the accurate selection of heart shadow from x-rays.

Methods and materials:

We collected a dataset of 1300 chest x-rays and applied masks to them. For analysis, all x-rays were converted to jpg, and the mask was created in png. To isolate the heart, we used a 9-layer U-net convolutional neural network. With startup parameters learning rate = 1e-5, Adam optimiser, batch = 5 (3, 5), epochs = 800. The network was launched with samples of 400, 800 and 1300 images. In each run, a generator of additional images was used (5 to 1 image every epoch).

Results:

The direct relationship is determined: an increase in the quality of segmentation with an increase in the number of x-rays in the dataset on the training sample. The quality of the test data is also affected by significant parameters when selecting lr, batch size. The parameter of epochs=800 shows good convergence on the graphs and was chosen as the most optimal. Samples with 400 and 600 images contain mask artefacts in the inner contour of the heart shadow. When sampling 1300 images, there are artefacts of determination beyond the outer contour.

Conclusion:

The accuracy of the contours (the pulmonary artery and the left atrial abalone) and the presence of artefacts depend on the volume of the dataset. The selection of optimal hyperparameters also affects the results of segmentation in more global moments, such as determining the main fixation zones.

Limitations:

The lower border of the heart remains fuzzy, even with a large number of images.

Ethics committee approval

n/a

Funding:

No funding was received for this work.