A Novel Sub-Pixel-Shift-Based High-Resolution X-ray Wired Flat Panel Detector (MRD1717)

X-ray flat panel detector (FPD) is a photographic element used in digital radiography. In the same way that a normal digital camera uses a CMOS sensor to receive light passing through a lens and converts it into an image, a flat panel detector converts X-rays passing through the object into a digital image.

X-ray flat panel detectors have been widely used in security, industrial, and medical applications in place of conventional image intensifiers (I.I.) and imaging plates (IP). The dynamic range of FPD is greater than that of I.I., and the images can be viewed in real time, without the need to remove the plate and extract the images, as with IP. The main applications of FPD in the industrial field are X-ray and CT non-destructive inspection, including casting and welding inspection, 3D printing inspection, SMT (surface-mounted technology) and semiconductor inspection, new energy battery inspection, security check, and so on; in the medical field, the main applications cover almost all X-ray equipment, including DR (digital radiography), DRF (dynamic DR), DM (digital mammography), CBCT (dental CT), DSA (digital subtraction angiography), C-arm X-ray systems, and so on. The first flat-panel detector DR systems based on amorphous silicon and amorphous selenium were introduced in 1995. Subsequently, major medical imaging equipment companies conducted preliminary research on the technology. In the late 1990s, GE and Perkin Elmer in cooperation, Thales, Siemens, Philips in co-investment with Trixell, Varex, Canon Medical, and other companies developed amorphous silicon flat panel detectors. Around 2010, amorphous silicon flat panel detector technology further proliferated, and traditional film players Carestream, Fujifilm, Konica, Agfa and e-Join also developed flat panel detectors.

Several kinds of FPDs have been developed over the past few decades. Existing FPDs are divided into two types: indirect conversion detectors and direct conversion detectors. The principle of indirect conversion FPDs is that X-ray irradiation is first converted to visible light through the scintillator; then, the digital image is read out using the principle of visible light cameras. Its basic structure includes the following: scintillator, sensor and readout circuit, and peripheral control circuit. The scintillator and sensor are the core part and determine the main performance of the FPD. Amorphous silicon (a-Si), CMOS, IGZO, and flexible FPD are all indirect conversion detectors. In contrast, direct conversion FPDs do not require scintillators. They convert X-rays into electrical signals directly after the X-rays are collected by the photoconductive semiconductor material. Therefore, the basic structure of the direct conversion FPD includes the following: sensor and readout circuit, peripheral control circuit, and the sensor (photoconductive semiconductor), which is the core part.

An indirect detector contains a layer of scintillating material that converts the X-rays into visible photons (light). Behind the scintillator, an array of photodiodes converts the light into an electrical signal. The array of photodiode pixels is similar in concept to a camera’s image sensor: a high density of pixels creates a high-resolution image in which small features are clearly and sharply rendered. The stored charge of each pixel is proportional to the intensity of the incident X-rays. Under the action of the control circuit, the stored charge of each pixel is scanned and read out, and the digital signal is output after A/D conversion and transmitted to the computer for image processing to form a digital X-ray image. A based architecture demonstration of the indirect FPD is illustrated in Figure 1.

The technologies used by different types of detectors and their main application areas are shown in Table 1.

The two most important metrics to evaluate the imaging quality of FPDs are the signal-to-noise ratio (SNR) and spatial resolution (SR). FPDs have different SNR and SR due to different materials, structures, and processes. SNR affects the ability to distinguish density differences in different tissues (i.e., high SNR means high density resolution), and SR affects the ability to distinguish fine spatial structures of tissues. To improve the spatial resolution, the pixel size of the detector needs to be made smaller. However, too small a pixel size leads to a decrease in SNR. As a direct consequence, the density resolution decreases, the image signal-to-noise ratio deteriorates, and the image becomes unusable. It is generally necessary to find a balance between these two imaging indexes. There is no FPD on the market that performs well in both SNR and spatial resolution.

In the medical field, the target of DR equipment is to observe and distinguish the density of different tissues in the chest and lungs. Therefore, the requirements for density resolution (SNR) are relatively high, while the pixel size of FPDs are larger, generally 139 μm, to easily obtain images with higher contrast. Accordingly, amorphous silicon FPDs are generally selected, which are more conducive to diagnosis. For the examination of extremity joints and breast, better imaging of structural spatial details is needed, so the pixel size of the FPD should be small, generally 50 to 76 μm, in order to obtain high-spatial-resolution images. Generally, amorphous selenium or CMOS FPDs are selected for these types of applications. The main reason for the higher requirement of spatial resolution in the industrial field than in the medical field is the continuous improvement of product quality control requirements. Defects that need to be detected are getting smaller and smaller as the result of the process upgrading. At present, film imaging is still used in a large number of industrial inspection scenarios, mostly on weld and casting inspection. In addition to regulatory factors, the main reason why users choose film over FDPs is that the actual spatial resolution of FPDs cannot reach the film imaging level. The industry expects that the spatial resolution could be further improved without decreasing SNR.

The resolution of the original image can be improved by pure software methods, which some people call super-resolution. Currently, there are three methods to achieve image super-resolution by software: interpolation-based methods, reconstruction-based methods, and learning-based methods.

Interpolation-based methods try to increase image resolution by filling in the corresponding pixel values on the empty spots after zooming in. However, these methods are efficient but ineffective. Reconstruction-based methods align multiple low- resolution images of the same scene with sub-pixel accuracy on the space, obtaining the motion offsets between high- and low-resolution images and constructing the spatial motion parameters in the observation model to obtain a high-resolution image. The core idea of reconstruction-based super-resolution methods is to trade temporal bandwidth (acquiring multiple image sequences of the same scene) for spatial resolution. With the significant development of deep learning technology, image super-resolution technology does have a wide range of practical applications in the fields of games, movies, medical imaging, and so on. Learning-based methods adopt the end-to-end mapping function of low-resolution images to high-resolution images by neural networks. Using the prior knowledge acquired by the model to obtain high-frequency details of the image, they are considered the best way to enhance image resolution at present. The mainstream algorithms are SRCNN, SRGAN, ESRGAN, and so on. Among them, the SRCNN method has the simplest network structure, using only three convolutional layers, and the framework is flexible in choosing parameters and supports customization. The disadvantage is that the details are not sufficiently represented, and the results obtained are too smooth and un- realistic when the magnification exceeds four. The SRGAN method generates realistic textures for a single image to complement the lost details but introduces high-frequency noise at the same time. The ESRGAN output image has better image quality with more realistic and natural textures. In the experimental section, we will cite the results of ESRGAN as a part of the comparison data.

The spatial resolution of FPDs is mainly determined by the pixel size, and a smaller pixel size leads to a higher spatial resolution. However, as a result of various factors such as process, cost, imaging quality, and imaging field of view, the pixel size cannot be reduced indefinitely at the actual product level. The common pixel size of amorphous silicon detectors is generally 100, 139, and 200 μm. The actual measured spatial resolution is worse than the theoretical spatial resolution corresponding to the pixel size. Types of scintillators (CSI, GdOS, etc.), scintillator thickness, and vapor deposition process will affect the actual spatial resolution. The measured spatial resolution of the detector with a nominal pixel size of 139 μm is typically between 150 to 190 μm.

Due to process and other constraints, the pixel size of the detector panel is limited. We have developed a novel high-resolution X-ray FPD based on sub-pixel shift (SPS) technology, which can improve the inherent spatial resolution of the detector without degrading the image quality.

The entire architecture of the proposed novel detector is shown in Figure 2, using CSI scintillator as the sensor; a-Si-based TFT as the converter; data operation part with integrated read, select, and pack; and a main state machine to control multiple modules and operation modes. Command passing through cmd process, buffer, and synchronization are designed. Udp/ip protocol stack and 1000 M wired network are used for all the command and data transfer.

A circuit design diagram of some basic hardware modules of the detector is shown in Figure 3, which includes: (a) AD chip interface, (b) gate driver chip interface, (c) network interface, and (d) main control board.

We use a sub-pixel shift design that allows each pixel to be spatially displaced in a controlled and precise sub-pixel scale in the XY direction, thus realizing pixel interpolation at the physical hardware level to improve the spatial resolution of the detector. We designed and implemented the underlying hardware and software for sub-pixel acquisition based on an amorphous silicon glass substrate, and users can easily control the sub-pixel acquisition accuracy of the detector through commands. Experiments show that the measured pixel size of this new high-resolution detector (MRD1717) is reduced from 162 to 132 μm, that is, the basic spatial resolution of the detector is improved by 18.5%.