Applications of convolutional neural networks in x-ray imaging

Researchers at the University of Canberra are looking for clinical experts who maybe seeking to undertake a Master by Research or PhD. Details of the project are:

1. Introduction

Visual perceptions begin at the moment light meets the retina. Retina consists of a sheet of photoreceptors that convert light to electrical signals. These signals are sent via the optic nerve to the primary visual cortex (figure 1(a)). Signal transmissions of visual perceptions can be emulated in computer science, such as deep neural network algorithms that imitate neural processes. In computer science, neural networks which have each neuron in one layer is connected to all neurons in the next layer are fully connected (figure 1(b)) and known as multiplayer perceptrons (Gardner & Dorling, 1998). Convolutional neural networks (CNNs) are a subset of deep neural networks which can be applied to visual image analysis.

CNNs emulate animal visual processes in which an individual visual-cortex neuron reacts to stimuli only in a restricted region of the entire visual field (Fukushima, 1980; Matsugu, Mori, Mitari, & Kaneda, 2003). This restricted region is known as the receptive field of the neuron, and the receptive fields of nearby neurons partially overlap, resulting the entire visual field completely covered by the receptive fields of all neurons (Hubel & Wiesel, 1968). In applied computation and visual-imagery analysis, CNNs are a class of deep learning neural networks that are trained on existing data and then used to predict outcomes, e.g. voice recognition, facial recognition, language translation, or computer-aided diagnosis (CAD).

In medical imaging, an example application of CNNs is mammogram CAD by a number of works (Arevalo, González, Ramos-Pollán, Oliveira, & Lopez, 2016; Carneiro, Nascimento, & Bradley, 2015; Qiu et al., 2016; Shin et al., 2016; Wei, Li, & Huang, 2011).  The applications of CAD in medical imaging still require a human expert to confirm the diagnosis before it can be accepted. Human experts still have to look at the images and based on the information shown in the images, make a diagnosis.

Human experts rely on sufficient image quality to make confident diagnosis from images. X-ray image quality depends on multiple factors. The amount of radiation reaching the image detector have significant effects on the image quality. Within a certain radiation range, the higher the radiation is reaching the image detector, the better the image quality is the final image. Scatter radiation reaching the image detector is one of the main noise factors that degrade the image quality. When the relative amount of scatter radiation to primary radiation is large in the image, the image quality is often severely degraded and interferes human experts making confident diagnosis.

2. Preliminary findings

Preliminary work has been focusing on devising a CNN algorithm for x-ray image processing. The preliminary evaluation of a CNN algorithm has already been completed. The CNN algorithm is illustrated in figure 2 in which multiple features are extracted from the original radiograph and then used to enhance it to obtain an outcome of enhanced image quality. This CNN algorithm has been evaluated with both a simulated image and an abdomen radiograph.

In the simulation evaluation of this CNN algorithm, a PMMA phantom (30 cm x 30 cm x 30 cm) with two disc-shaped objects were used to imitate a large abdomen radiographic condition and the radiation exposures were set at one-thousandth of a typical radiation exposure. The original simulated image (3.a) fails to show the projections of the two objects. After CNN enhancement, the two projections are clearly visible (3.b). This might mean that with a perfect image receptor like the one used in the simulation, radiation exposures could be reduced by thousand folds while the image quality would still be good enough for making confident diagnosis.

In the evaluation of this CNN algorithm using an abdomen radiograph, the abdomen radiograph was taken at 75 kVp and 16 mAs without using an anti-scatter grid. The abdomen size was approximately 25 cm thick. The exposure for such abdomen sizes is about 75 kVp and 40 mAs and with an anti-scatter grid. The original radiograph is shown in figure 4.a and is not of diagnostic quality. Human experts would have low confidence in making a diagnosis from this radiograph. After enhancing it by the CNN algorithm, the enhanced outcome is shown in figure 4.b which shows significant improvement in image quality and human experts would be confidently making diagnosis from this radiograph. This means the CNN algorithm has the potential to reduce radiation exposures to patients, in this example exposure reduction by 60%, while still achieving high-quality useful radiographs.

3. Research project

The aims of this research project are to assist human experts making confident diagnosis and at the same time reduce radiation exposures to patients hence reduce radiation-exposure-induced cancer risks for patients undergoing x-ray examinations. This project aims at two objectives: one is enhancing image quality and reducing radiation exposures to patients, the other is achieving CAD. This project may carry out in two stages. In the first stage, we will use a CNN algorithm to enhance underexposed images and provide sufficiently high image quality to assist human experts making confident diagnosis. In the second stage, the CNN algorithm will be revised and applied to make diagnosis.

3.1 Image quality enhancement using CNNs

The first stage of this research project is to evaluate the CNN algorithm for image quality enhancement. Image quality may be determined statistically or perceptually. Statistical image quality assessments are such as contrast-detail analysis (Geijer & Persliden, 2005; Perez-Ponce, Daul, Wolf, & Noel, 2011), signal-noise-ratio analysis (CNR) (Huda, Sajewicz, Ogden, & Dance, 2003), figure of merit (FOM) analysis (Fausto et al., 2017). Perceptual image quality evaluation methods are such as visual grading characteristic (VGC) analysis (Båth & Månsson, 2007; Ludewig, Richter, & Frame, 2010), visual grading analysis (VGA) (Sund, Båth, Kheddache, & Månsson, 2004; Ullman et al., 2006). The CNN algorithm is ideally evaluated using either VGC or VGA, in additional to that, contrast-detail analysis is also an added credit.

Another useful evaluation for this CNN algorithm is a cross-comparison evaluation with other image quality enhancement methods, such as the Contrast Limited Adaptive Histogram Equalization (CLAHE) (Koonsanit, Thongvigitmanee, Pongnapang, & Thajchayapong, 2017), the Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) (Sheet, Garud, Suveer, Mahadevappa, & Chatterjee, 2010), and the Histogram Equalization (HE) (Roopaei, Agaian, Shadaram, & Hurtado, 2014). A example of similar comparison publications is the study of Saenpaen, Arwatchananukul, and Aunsri (2018).

3.2 CNN algorithm for diagnosis

To apply this CNN algorithm to diagnosis may be continued after the first stage is completed or it may be concurrently performed with the image quality evaluation. Nonetheless it should be separated from the image quality evaluation. Some CNN applications in medical x-ray image are mammogram diagnosis (Arevalo et al., 2016; Carneiro et al., 2015; Qiu et al., 2016; Wei et al., 2011), lung disease diagnosis (Cheng, Feng, & Jia, 2018), vessel extraction in x-ray angiograms (Nasr-Esfahani et al., 2016).

4. Research candidates

The research activities proposed above are suitable for both research master and PhD candidates.

The research candidates are responsible for image quality evaluation activities by VGC or VGA as well as other image quality evaluation methods. The candidates under the supervision will be expected to perform a variety of research activities, which include literature review, ethic applications, recruitment of participants (if needed), data analysis and dissemination in peer reviewed journals, as well as other research activities within the overall scope of a research project.

PhD candidates who are interested in expanding applications of this CNN algorithm are also welcome to join the team.

Contacts:

Principle Supervisor:

Dr Abel Zhou, PhD.  Abel.Zhou@canberra.edu.au

Co-supervisors:

Prof Rob Davidson, PhD.  Rob.Davidson@canberra.edu.au

Prof Nick Brown, PhD.   Nick.Brown@canberra.edu.au