Adam Pantanowitz

Bradley Segal, David M. Rubin, Grace Rubin et al.

Chest x-rays are a vital tool in the workup of many patients. Similar to most medical imaging modalities, they are profoundly multi-modal and are capable of visualising a variety of combinations of conditions. There is an ever pressing need for greater quantities of labelled data to develop new diagnostic tools, however this is in direct opposition to concerns regarding patient confidentiality which constrains access through permission requests and ethics approvals. Previous work has sought to address these concerns by creating class-specific GANs that synthesise images to augment training data. These approaches cannot be scaled as they introduce computational trade offs between model size and class number which places fixed limits on the quality that such generates can achieve. We address this concern by introducing latent class optimisation which enables efficient, multi-modal sampling from a GAN and with which we synthesise a large archive of labelled generates. We apply a PGGAN to the task of unsupervised x-ray synthesis and have radiologists evaluate the clinical realism of the resultant samples. We provide an in depth review of the properties of varying pathologies seen on generates as well as an overview of the extent of disease diversity captured by the model. We validate the application of the Fréchet Inception Distance (FID) to measure the quality of x-ray generates and find that they are similar to other high resolution tasks. We quantify x-ray clinical realism by asking radiologists to distinguish between real and fake scans and find that generates are more likely to be classed as real than by chance, but there is still progress required to achieve true realism. We confirm these findings by evaluating synthetic classification model performance on real scans. We conclude by discussing the limitations of PGGAN generates and how to achieve controllable, realistic generates.

Adam Pantanowitz, Kimoon Kim, Chelsey Chewins et al.

A custom head-mounted system to track smooth eye movements for control of a mouse cursor is implemented and evaluated. The system comprises a head-mounted infrared camera, an infrared light source, and a computer. Software-based image processing techniques, implemented in Microsoft Visual Studio, OpenCV, and Pupil, detect the pupil position and direction of pupil movement in near real-time. The identified direction is used to determine the desired positioning of the cursor, and the cursor moves towards the target. Two users participated in three tests to quantify the differences between incremental tracking of smooth eye movement resulting from the Vestibulo-ocular Reflex versus step-change tracking of saccadic eye movement. Tracking smooth eye movements was four times more accurate than tracking saccadic eye movements, with an average position resolution of 0.80 cm away from the target. In contrast, tracking saccadic eye movements was measured with an average position resolution of 3.21 cm. Using the incremental tracking of smooth eye movements, the user was able to place the cursor within a target as small as a 9 x 9 pixel square 90 % of the time. However, when using the step change tracking of saccadic eye movements, the user was unable to position the cursor within the 9 x 9 pixel target. The average time for the incremental tracking of smooth eye movements to track a target was 6.45 s, whereas for the step change tracking of saccadic eye movements, it was 2.61 s.

Adam Pantanowitz, Emmanuel Cohen, Philippe Gradidge et al.

Obesity is an important concern in public health, and Body Mass Index is one of the useful (and proliferant) measures. We use Convolutional Neural Networks to determine Body Mass Index from photographs in a study with 161 participants. Low data, a common problem in medicine, is addressed by reducing the information in the photographs by generating silhouette images. Results present with high correlation when tested on unseen data.

Steven Dinger, John Dickens, Adam Pantanowitz

Recent advances in telecommunications have enabled surgeons to operate remotely on patients with the use of robotics. The investigation and testing of remote surgery using a robotic arm is presented. The robotic arm is designed to have four degrees of freedom that track the surgeon's x, y, z positions and the rotation angle of the forearm θ. The system comprises two main subsystems viz. the detecting and actuating systems. The detection system uses infrared light-emitting diodes, a retroreflective bracelet and two infrared cameras which as a whole determine the coordinates of the surgeon's forearm. The actuation system, or robotic arm, is based on a lead screw mechanism which can obtain a maximum speed of 0.28 m/s with a 1.5 degree/step for the end-effector. The infrared detection and encoder resolutions are below 0.6 mm/pixel and 0.4 mm respectively, which ensures the robotic arm can operate precisely. The surgeon is able to monitor the patient with the use of a graphical user interface on the display computer. The lead screw system is modelled and compared to experimentation results. The system is controlled using a simple proportional-integrator (PI) control scheme which is implemented on a dSpace control unit. The control design results in a rise time of less than 0.5 s, a steady-state error of less than 1 mm and settling time of less than 1.4 s. The system accumulates, over an extended period of time, an error of approximately 4 mm due to inertial effects of the robotic arm. The results show promising system performance characteristics for a relatively inexpensive solution to a relatively advanced application.