External beam radiotherapy (EBRT) treatment planning requires a fast and accurate method of calculating the dose delivered by a clinical treatment plan. However, existing methods of calculating dose distributions have limitations. Monte Carlo (MC) methods are accurate but can take too long to be clinically viable. Deterministic approaches are quicker but can be inaccurate under certain conditions, particularly near heterogeneities and air interfaces. Neural networks trained on MCderived data have the potential to reproduce dose distributions that agree closely with the MC method while being significantly quicker to deploy. Methods: In this work we present a framework for training machine learning models capable of directly calculating the dose delivered to a point in three-dimensional (3D) heterogeneous media given only spatially local information. The framework consists of three parts. First, we describe a novel method of randomly generating 3D heterogeneous geometries using simplex noise. Dose distributions for training were obtained by importing these geometries into a MC simulation. The second and third parts of the framework are precalculated data channels, aligned with the patient computed tomography (CT) image, to be used as input to the model. These data channels are a computationally efficient way of encoding the parameters of an incident radiation beam while also allowing the model to learn from data that would otherwise be outside of its receptive field. Results: We demonstrate the viability of the framework by a training small, fully connected neural network model to reproduce dose distributions from megavoltage photon beams. The trained network displayed excellent agreement with MC dose distributions in randomly generated geometries with an average gamma index (3%/3 mm) pass rate of 94.7% and an average error of 1.45% of peak dose. Finally, the network was used to calculate the dose in a patient CT image, on which the network was not trained, producing similarly impressive results. Conclusions: A novel method of generating training data for learned radiation dosimetry models has been introduced, along with preprocessing steps that allow even simple models to reproduce accurate dose distributions for EBRT. More importantly, we have demonstrated that a model trained using the proposed framework can generalize from the training data to predicting the therapeutic dose in realistic media.

No abstract

Prolonged sitting and inadequate sleep can impact driving performance. Therefore, objective knowledge of a driver’s recent sitting and sleep history could help reduce safety risks. This study aimed to apply deep learning to raw accelerometry data collected during a simulated driving task to classify recent sitting and sleep history. Participants (n = 84, Mean ± SD age = 23.5 ± 4.8, 49% Female) completed a seven-day laboratory study. Raw accelerometry data were collected from a thigh-worn accelerometer during a 20-min simulated drive (8:10 h and 17:30 h each day). Two convolutional neural networks (CNNs; ResNet-18 and DixonNet) were trained to classify accelerometry data into four classes (sitting or breaking up sitting and 9-h or 5-h sleep). Accuracy was determined using five-fold cross-validation. ResNet-18 produced higher accuracy scores: 88.6 ± 1.3% for activity (compared to 77.2 ± 2.6% from DixonNet) and 88.6 ± 1.1% for sleep history (compared to 75.2 ± 2.6% from DixonNet). Class activation mapping revealed distinct patterns of movement and postural changes between classes. Findings demonstrate the suitability of CNNs in classifying sitting and sleep history using thigh-worn accelerometer data collected during a simulated drive. This approach has implications for the identification of drivers at risk of fatigue-related impairment.

Radiation therapy requires clinical linear accelerators to be mechanically and dosimetrically calibrated to a high standard. One important quality assurance test is the Winston-Lutz test which localizes the radiation isocentre of the linac. In the current work we demonstrate a novel method of analysing EPID based Winston-Lutz QA images using a deep learning model trained only on synthetic image data.In addition, we propose a novel method of generating the synthetic WL images and associated 'ground-truth' masks using an optical ray-tracing engine to 'fake' mega-voltage EPID images. The model called DeepWL was trained on 1500 synthetic WL images using data augmentation techniques for 180 epochs. The model was built using Keras with a TensorFlow backend on an Intel Core i5-6500T CPU and trained in approximately 15 hours. DeepWL was shown to produce ball bearing and multi-leaf collimator field segmentations with a mean dice coefficient of 0.964 and 0.994 respectively on previously unseen synthetic testing data. When DeepWL was applied to WL data measured on an EPID, the predicted mean displacements were shown to be statistically similar to the Canny Edge detection method. However, the DeepWL predictions for the ball bearing locations were shown to correlate better with manual annotations compared with the Canny edge detection algorithm. DeepWL was demonstrated to analyse Winston-Lutz images with accuracy suitable for routine linac quality assurance with some statistical evidence that it may outperform Canny Edge detection methods in terms of segmentation robustness and the resultant displacement predictions.

In recent years, machine learning approaches have been used to classify and extract style from media and have been used to reinforce known chronologies from classical art history. In this work we employ the first ever machine learning analysis of Australian rock art using a data efficient transfer learning approach to identify features suitable for distinguishing styles of rock art. These features are evaluated in a one-shot learning setting. Results demonstrate that known Arnhem Land Rock art styles can be resolved without knowledge of prior groupings. We then analyse the activation space of learned features and report on the relationships between styles and arrange these classes into a stylistic chronology based on distance within the activation space. By generating a stylistic chronology, it is shown that the model is sensitive to both temporal and spatial patterns in the distribution of rock art in the Arnhem Land Plateau region. More broadly, this approach is ideally suited to evaluating style within any material culture assemblage and overcomes the common constraint of small training data sets in archaeological machine learning studies.