Using deep learning for pixel-defect corrections in flat-panel radiography imaging

Abstract.

Purpose: Flat-panel radiography detectors employ thin-film transistor (TFT) panels to acquire high-quality x-ray images. Pixel defects occur due to circuit shorts or opens in the TFT panel. The defects may degrade the image quality, as well as lower the production yield, and eventually raise the production cost. Hence, it is important to develop an appropriate defect correction algorithm for acquired images. Traditional correction algorithms are based on a complicated adaptive filtering technique, which exploits neighbor pixels, to faithfully preserve the edge components. Because of the complexity of the traditional sophisticated approaches, optimizing their correction performances is difficult.

Approach: We considered various pixel-defect correction algorithms based on different deep learning models, such as the artificial neural network (ANN), convolutional neural network (CNN), concatenate CNN, and generative adversarial networks (GAN). We considered two cases of maximal defect sizes, $3\times 3$ and $5\times 5\text{pixels}$, and conducted extensive learning experiments to find the best structures of the learning models using the mean square error (MSE) as the loss function.

Results: To conduct experiments, practical chest x-ray images were acquired from a general radiography detector. The MSE values of the correction results from ANN, CNN, concatenate CNN, and GAN were 69.40, 75.13, 68.21, and 73.77, respectively, and were much smaller than that of the conventional template match correction method.

Conclusions: A concatenate CNN showed the best defect-correction performance. However, ANN could achieve a similar correction performance with much smaller encoding complexity. Therefore, the single-layer ANN can efficiently conduct defect corrections in terms of both correction and complexity.

1. Introduction

Research on digital radiography (DR) technology is actively advancing in medical imaging. The DR technology can be implemented based on flat-panel radiography detectors.^1–3 In an indirect flat-panel detector, charges are generated from photodiodes, which can sense the generated light photons from a scintillator layer. Here, the photodiodes are controlled by a thin-film transistor (TFT) array. In a direct flat-panel detector, charges are generated through interactions of x-ray photons with a photoconductor layer and are read out through charge collectors, which are also controlled by a TFT array. As shown in Fig. 1(a), the generated charges are stored in the capacitors or photodiodes of the TFT panel and are converted into voltage signals through the charge amplifiers of the readout integrated circuit. A digital image is finally extracted through the analog-to-digital converters.²

Fig. 1.

Flaws in TFT flat-panel detectors: (a) structure of a direct TFT flat-panel detector, staining in the photoconductor and pixel defects and (b) structure of a pixel in an indirect detector. Electrical damage can occur at the overlaps and electrodes of the data, gate, and bias lines, and the TFT and photodiode electrodes as open or short circuits.

In the flat-panel detectors, staining of the photoconductor or scintillator above the TFT array causes abnormal responses to the corresponding pixels as shown in Fig. 1(a). Flaws in the TFT array generally occur due to electrical circuit damages at the time of manufacture, including the TFT process.⁴ Among the flaws, pixel defects due to electrical circuit damages in the TFT array can be classified into the following types: point, cluster, and line defects as shown in Fig. 2. The point defect implies a single defective pixel and the cluster defect implies two or more defective pixels attached to each other. The line defect implies the defective pixels that show a line shape. The causes of such defects are electrical open and short circuits, including current leakage due to poorly completed circuits,⁶ or cracks in the insulating film.⁷ Open circuits can occur at the gate lines, data lines, bias lines, and TFT electrodes.⁸ Incomplete via holes can also open the circuits. The remaining pieces of insulating film during the TFT process may cause short circuits between the source and drain electrodes of field-effect transistors.⁹ Short circuits can occur at overlaps and electrodes between wiring lines, such as gate lines, data lines, bias lines, and TFT electrodes, as shown in Fig. 1(b).⁸ The storage capacitors can also be shorted. These flaws cannot faithfully produce the corresponding gray levels and can yield white or black levels. Thus, the flaws can seriously degrade the quality of the acquired x-ray images¹⁰ as shown in Fig. 2. To increase the production yield, and thus reduce the manufacturing cost,¹¹ it is necessary to perform detection or classification,^6,12–15 and then correction of the defective pixels for the acquired images.^16–21

Fig. 2.

Types of defective pixels in a flat-panel radiography detector (DRTECH Co. Ltd.⁵). (a) Point defect, (b) cluster defect, and (c) line defect.

In developing TFT arrays, we can carefully design circuit structures that are resilient to circuit damages.^7,22 The flaws caused from electrical leak over the TFT array, or lines and pixel electrodes can be corrected by performing laser trimming operations.² However, it is quite difficult to detect flaws in an active matrix substrate before assembly, which thus leads to an increase in costs. Hence, practically, it is difficult to correct defective pixels if the flat-panel detector has a large number of pixels. Because of the high cost of trimming the TFT array, first acquiring digital images and then correcting the values of defective pixels is more practical and efficient than trimming.^2,4,23 After defective pixels are detected, the bad pixel values can be corrected by another values. The correction should conceal the defects of the acquired image and minimize producing artifacts after image processing.²⁴ Here, the correction values can be generated from an average of the nearest-neighbors that are not defective.²⁵

For the single point defect case of Fig. 2(a), various correction algorithms for defective pixels have been developed.^16,26–28 The median filter can efficiently correct the defective single pixel using its impulsive property.^2,29 A linear interpolation using neighbor pixels can also be used.²⁶ To preserve an edge pattern along the defective pixel, the surrounding 6 or 8 pixels can be used under sophisticated pattern detection schemes in the spatial domain.^16,28 A domain of the discrete cosine transform can also be considered to describe edges.³⁰ For the cluster defect case of Fig. 2(b), which has more than a single defective pixel, the median filter does not work. Several correction kernels can be considered for the cluster defect and corrections can be conducted according to pattern types in a similar manner to the single point defect case.³¹ A template matching correlation (TMC) technique can be used to restore edge patterns in a defective cluster $3\times 3\text{pixels}$.³² However, the TMC technique requires a search region that is much larger than that of the defective pixels. Thus, the implementation complexity is quite high. A spectral deconvolution based on the Fourier transform can be used to restore the defective cluster pixels and can alleviate defects of large areas by observing the spectrum of the whole image.³³ However, iterative deconvolutions require large memories, high computational complexities, and fast execution times. Furthermore, extensive simulations show that considering the whole image property is not efficient in correcting the cluster defect compared to locally considering neighbor pixels. Hence, this approach limits its application to practical pixel-defect corrections for a purpose of hardware-based fast correction schemes. Recent research, which has been actively conducted on deep learning, has shown good results, especially in the area of inpainting.^34–36 Hence, such inpainting approaches can be candidates in developing algorithms to correct the cluster defect.

For the detection and classification of defective pixels, we can find numerous research studies that employ the deep learning technique in the literature.^37–39 However, it seems that deep learning techniques are not employed in pixel-defect corrections especially for the cluster defect. If we consider the cluster defect rather than the point defect, then further efficient techniques, such as the deep learning technique, should be considered. In this paper, in order to correct the cluster defect of Fig. 2(b), in which their maximal size is $3\times 3\text{pixels}$, we consider several deep learning models, such as the single-layer artificial neural network (ANN), convolutional neural network (CNN), concatenate CNN, and generative adversarial networks (GAN) models, using the mean square error (MSE) as the loss function.⁴⁰ The corrections are then extended to the cases of $5\times 5\text{pixels}$. The pixel-defect correction algorithms, which are based on deep learning techniques, can offer advantages that the required pixels to correct the pixel defect can be minimized, compared to the conventional TMC technique case with relatively low computational complexities.

The paper is organized as follows. Section 2 introduces pixel-defect correction algorithms based on the single-layer ANN, CNN, concatenate CNN, and GAN models, respectively. Section 3 summarizes the experimental results with comparisons, whereas Sec. 4 concludes the paper.

2. Pixel-Defect Corrections Based on Deep Learning

In this section, we introduce several pixel-defect correction algorithms that are based on deep learning. For the correction in this paper, we consider the cluster defect and assume that the maximal defect size is $3\times 3$ or $5\times 5\text{pixels}$.

For the case of the maximal $3\times 3\text{pixels}$, Fig. 3 summarizes procedures of pixel-defect corrections. We assume that the positions of the defective pixels are known, and the defective block includes them. In the example of Fig. 3, the defective block size is $3\times 3\text{pixels}$. We then consider an image block of $7\times 7\text{pixels}$, in which the center part is the corresponding defective block, as shown in Fig. 3. The correction algorithm estimates all $3\times 3\text{pixels}$ of the defective block using the pixels of the image block, excluding the defective block. Here, a neural network is trained using image blocks without any defects, by using the center pixels that correspond to the defective block as reference output values and the neighbor pixels of the image block as the input values for supervised learning. As the training phase continues, the outputs gradually resemble the reference values and can be used as an appropriate estimate of the defective block of $3\times 3\text{pixels}$. After finishing the training phase, the weights will be used to correct the pixel defects. Among the estimated pixels of the defective block, we use only the values corresponding to the defective pixels while maintaining the normal pixels in the defective block.

Fig. 3.

Structures for pixel-defect corrections based on deep learning. The algorithm estimates the center defective block of $3\times 3\text{pixels}$, including the defective pixels. The defective pixels are substituted by the corresponding estimated pixels: (a) using the surrounding $40(=7\times 7-3\times 3)\text{pixels}$ outside the defective block and (b) using $49(=7\times 7)\text{pixels}$, including roughly corrected defective pixels.

For the structure of Fig. 3(a), the neighbor pixels of the defective cluster pixels within the defective block are not used to estimate the defective pixels. On the other hand, under deep learning models, those neighbor pixels can easily be utilized to estimate the defective pixels, as shown in Fig. 3(b). Here, the defective pixels can first be roughly corrected by using an average of neighbor pixels, and then an image block including the roughly corrected pixels can enter a deep learning model. For a simple comparison of various deep learning models, we consider the structure of Fig. 3(a) in this paper.

2.1. Single-Layer Artificial Neural Network

The values for the defective block can be directly estimated based on a simple ANN model. Among the ANN models, a single-layer ANN, which has no hidden layers, can be successfully employed to correct the defective pixels, because of its efficient estimate with low computational complexity, especially for a small defective block of $3\times 3$ or $5\times 5\text{pixels}$. Figure 4 shows a single-layer ANN, which is used to correct defective pixels. For the estimate of the $5\times 5$ defective block, $56(=9\times 9-5\times 5)\text{pixels}$, which surround the defective block, the 56 nodes of the input layer are entered in the fully connected (FC) network. The output layer of the single-layer ANN then receives the weighted and biased inputs from the input layer to estimate the $5\times 5$ defective block. Next, the defective pixels are corrected by using only the corresponding pixel values from the estimated $5\times 5\text{pixels}$. In Fig. 5, the MSE values of the ANN model with a single layer are illustrated for different input sizes. For the cluster defects of $3\times 3$ and $5\times 5\text{pixels}$, optimal input sizes are $7\times 7$ and $9\times 9$, respectively, as shown in the ANN model of Fig. 4. We may add hidden layers to further improve the estimate performance. For the $3\times 3$ cluster defect case, adding a hidden layer even increased the MSE value from 64.0 to 64.4. Furthermore, adding another hidden layer could not decrease the MSE value as 64.1. A similar result could be obtained for the case of $5\times 5$ cluster defect. Through extensive experiments, we concluded that additional hidden layers did not improve the estimate performance. We can also consider the structure of Fig. 3(b) for this simple model. However, the simulation result of this structure was quite similar to that of Fig. 3(a).

Fig. 4.

Single-layer ANN model for the pixel-defect correction.

Fig. 5.

Parameter tuning example on ANN and CNN. The ANN model has a single layer. In the CNN model, the first layer has 8 filters with a size of $3\times 3$ and the second layer has 4 filters with a size of $3\times 3$.

2.2. Convolutional Neural Network

The single-layer ANN model flattens two-dimensional image blocks into one-dimensional data. Hence, the training phase for the single-layer ANN may not efficiently consider the two-dimensional spatial features. In other words, for the single-layer ANN model, it is difficult to grasp spatial features of an image block and learn it efficiently. On the other hand, the CNN model can be learnt through the convolutional operation, whereas maintaining the spatial features of an image block. Hence, the CNN model is widely used and is known to have superior performance, especially in the field of image recognition.⁴¹ The CNN model usually contains various combinations of convolutional and pooling layers that depend on applications to improve its performance. The convolution layer applies weights and biases to the inputs, and the pooling layer operates to select values from the previous convolutional layer after activation functions. Figure 6 shows a CNN model for the pixel-defect correction. A single-layer ANN model, such as that of Fig. 4, first estimates a coarse $5\times 5$ defective block. This coarse estimate is then inserted into the center part of a $15\times 15$ image block, as shown in Fig. 6. This modified $15\times 15$ image block then enters the main CNN part, and a fine estimate is finally produced through several convolutional layers. The main CNN part consists of two convolutional layers, as shown in Fig. 6. Here, each layer performs the convolution, and then batch normalization (BN) is used to yield outputs for the activation function.⁴² Here, the exponential linear unit (ELU) is employed as an activation function. In Fig. 5, the MSE values of the CNN model are also illustrated for different input sizes. For the case of the CNN model, we can observe that the MSE values decrease as the input size increases. Considering the computational complexity, we fixed the input size as $15\times 15$ as shown in Fig. 6. Through extensive hyperparameter tuning, we could notice that 16 or 8 filters with a convolutional kernel of $3\times 3$ for two convolutional layers were appropriate. Here, we may use the pooling layer after the convolutional layer. However, further improvements were not achieved.

Fig. 6.

CNN model, in which the upper single-layer ANN provides a coarse estimate for the lower main CNN stage. Each convolutional layer uses the ELU activation function after a BN process.

The conventional CNN models do not use the FC layer if they estimate relatively large image blocks. However, for the pixel-defect corrections, because restoring $3\times 3\text{pixels}$ can be regarded as a data reducing process and, in such a case, adding the FC layer can provide a better performance. From extensive experiments, we found that adding the FC layer yielded better performances than the case of the model without the FC layer.

2.3. Concatenate Convolutional Neural Network

From experiments, the maximal defect sizes, which were considered in this paper, were relatively small, and thus gains from the spatial features of CNN were not expected. To exploit the advantages of both the single-layer ANN model of Fig. 4 and the CNN model of Fig. 6, we can consider a concatenate CNN model, as shown in Fig. 7. Figure 7 shows that the concatenate CNN model combines the outputs of the single-layer ANN and CNN models through a concatenation layer. From extensive experiments for the $5\times 5$ defective block, we could notice that relatively small block sizes, e.g., $7\times 7$, were appropriate for the single-layer ANN, whereas relatively large block sizes, e.g., $15\times 15$, were appropriate for efficient capture of the two-dimensional spatial features. A coarse estimate of the $5\times 5$ defective block from the single-layer ANN is also used for the CNN stage as an input modification, in a similar way to the CNN model of Fig. 6. The concatenate layer in the right-hand side of Fig. 7 combines $993(=25+968)$ features, and the following FC layer yields a fine $5\times 5$ estimate for the pixel-defect correction.

Fig. 7.

Concatenate CNN model. The upper single-layer ANN of a relatively small input block is concatenated with the lower CNN of a relatively large input block.

2.4. Generative Adversarial Networks

The GAN model⁴³ can be successfully used to correct the defective pixels from an aspect of inpainting. For the pixel-defect correction, the least squares GAN model with the MSE loss function generates defective blocks, which are similar to the reference image block, based on a supervised learning for the two adversarial networks; generator and discriminator, as shown in Fig. 8. The generator uses a two-layer ANN model with a hidden layer and receives image blocks of $15\times 15\text{pixels}$ for generating the defective blocks of $5\times 5\text{pixels}$. Here, the hidden and output layers use the hyperbolic tangent (tanh) function as an activation function. The discriminator also uses a two-layer ANN model with $5\times 5$ inputs and 1 output. Here, the rectified linear unit and sigmoid functions are used for the hidden and output layers, respectively.

Fig. 8.

The GAN model, in which both the generator and discriminator employ the two-layer ANN model. The trained generator is used as an estimate algorithm.

3. Experimental Results

In this section, the pixel-defect correction results from the algorithms, which are based on various deep learning models, the single-layer ANN, CNN, concatenate CNN, and GAN, are compared with the conventional TMC algorithm. To conduct experiments, chest x-ray images were acquired from a general radiography detector, as summarized in Table 1 (DRTECH Co. Ltd.⁵), using a multipurpose chest phantom, N1, which is called “LUNGMAN,” Koyto Kagaku Co. Ltd. Here, we assume that the main application of the considered detector is for acquiring chest x-ray images. An acquired x-ray image has $3072\times 3072\text{pixels}$ with the gray-level resolution of 14 bits/pixel, and the pixel pitch is $140\mu \mathrm{m}/\text{pixel}$. The x-ray images were acquired under the RQA5 condition of the IEC62220-1-1 standard.⁴⁴ The chest x-ray images were cut into sequential image blocks by overlapping the chest portion, and these blocks were used as the inputs of the pixel-defect correction algorithms. Here, the image block size was dependent on the considered maximal defect size. We used about 176,800 training image blocks, 176,800 validation image blocks, and 116,000 testing image blocks from 10 chest x-ray images acquired at different exposures. We considered sizes of $3\times 3$ and $5\times 5\text{pixels}$ for the cluster defect in the experiments. To conduct deep learning, we used a workstation with NVIDIA Quadro P6000 under a framework of TensorFlow and Keras. The experiments were carried out with a batch size of 1000 and 1000 epochs. For the GAN case, the model was trained with exceptional 1,000,000 epochs to stabilize the results. In training the neural networks, we used MSE as the loss function. The MSE loss function can faithfully correct their mean and smooth varying gray levels rather than complicate structures within such a small $3\times 3$ block. We can also consider the structural similarity index measure (SSIM) as a loss function based on a perceived quality. For the SSIM case, we should consider convergence problems and develop appropriate variations of SSIM for the pixel-defect correction purpose.^45–47 Hence, considering SSIM can be a separate future work for the pixel-defect correction.

Table 1.

Development prototype of flat-panel general radiography detector (DRTECH Co. Ltd.).

Image size	$3072\times 3072\text{pixels}$
Collection element	a-Si TFT/photodiode
Scintillator	Columnar CsI (Tl)
Pixel pitch ($\mu \mathrm{m}/\text{pixel}$)	140
Gray level (bits/pixel)	14

To compare the correction algorithms of deep learning, a conventional TMC algorithm was considered, as shown in Fig. 9. In the TMC algorithm, first, the defective pixels are roughly corrected based on a simple averaging technique, by using neighbor normal pixels of the image block of $9\times 9\text{pixels}$ if the defective block size is $3\times 3$. Using the roughly corrected image block of $9\times 9\text{pixels}$, template matching is then conducted around a larger image block of $27\times 27\text{pixels}$. Finally, a fine correction is conducted using the best-matching image block to faithfully represent edge directions. The first averaging step can be implemented within an a field programmable gate array (FPGA) chip, whereas the second template matching is usually executed in a separate computer with ease. It is noteworthy that the computational complexity of the single-layered ANN can be an extended amount from that of the first TMC step, where a certain number of operations of multiplications and then adds are required. Hence, the simple single-layer ANN can be implemented within an FPGA chip or an on-board digital signal processing algorithm and thus can provide a fast pixel-defect correction without additional processing in a separate computer.^16,26 The iterative convolutional method³³ was also considered for a comparison. However, because the correction performance of this method was quite worse than the TMC case, we did not show the results from the iterative convolutional method in this paper.

Fig. 9.

Conventional TMC algorithm for the pixel-defect correction. For the case of the $3\times 3$ defect size, $27\times 27\text{pixels}$ are required for template matching. The defective pixels are first corrected using the average of neighbor pixels of $9\times 9\text{pixels}$, and then a template matching is conducted for the search region of $27\times 27\text{pixels}$ using the template of $9\times 9\text{pixels}$.

Tables 2–4 summarize the comparison results. Table 2 briefly describes the structures of the deep learning models and TMC. Computational complexities of the considered models can be observed from the required parameters as summarized in Table 3. From Tables 2–3, we notice that, among the considered models, the single-layer ANN has the simplest structure.

Table 2.

Comparison of the deep learning models.

	TMC	Single-layer ANN	CNN	Concatenate CNN	GAN
Input	2D spatial	1D data	2D spatial	2D spatial	2D spatial
Structure	Template matching	FC layer	Convolutional layer/BN	Concatenate ANN and CNN	Two-layer ANN

Table 3.

Comparison of the number of parameters for deep learning.

Defective block	Single-layer ANN	CNN	Concatenate CNN	GAN
$3\times 3$	369	10,137	10,587	16,827
$5\times 5$	1425	26,266	26,891	20,171

Table 4.

Comparison of the MSE values per pixel.

Defective block	TMC	Single-layer ANN	CNN	Concatenate CNN	GAN
$3\times 3$	215.7	69.40	75.13	68.21	73.77
$5\times 5$	243.6	94.67	100.7	91.80	99.41

Table 4 summarizes a comparison of MSE per pixel for the cluster defect. We can observe that the algorithms based on deep learning show much smaller MSE values than the conventional TMC case for both defective block sizes $3\times 3$ and $5\times 5\text{pixels}$. Among the algorithms in Table 4, the concatenate CNN model shows the best correction result. Even though the maximal pixel-defect size increases to $5\times 5$ and thus the corresponding MSE values slightly increase, the algorithms based on deep learning show better correction results than the $3\times 3$ case of TMC, and thus can be employed to correct $5\times 5\text{pixel}$ defects. From the result of Table 4, we can also observe that the MSE result of the single-layer ANN is even better than that of the GAN model. Furthermore, the computational complexity of the single-layer ANN is much lower than that of GAN. Hence, for the case of simple and real-time implementations of the pixel-defect correction, employing the single-layer ANN can offer a great advantage. Similar MSE comparison analysis could be observed for different image sequences, which were acquired under the same RQA5 condition of the IEC62220-1-1 standard but different dose conditions.⁴⁰

To solve the correction problem for pixel defects that occur in the TFT flat panels of radiography detectors, we have conducted experiments and comparisons using conventional inpainting techniques based on CNN and GAN. However, from extensive experiments, we have found that the complicate inpainting techniques could not dramatically improve correction performances over the single-layer ANN. The problem for correcting pixel defects is restoring very small image blocks, in which the maximal size is as small as $3\times 3\text{pixels}$. It is noteworthy that the conventional inpainting techniques are focusing on larger image blocks. The experimental results of the paper thus indicate that corrections of small image blocks are not efficient for the cases of the conventional inpainting techniques. In other words, for the pixel-defect correction, using a simple structure of the single-layer ANN in Fig. 4 is enough to correct small image blocks of $3\times 3\text{pixels}$.

Figures 10 and 11 show examples of the corrected results. Here, we selected the image blocks in which the TMC algorithm failed badly to correct the defect. We can observe that the TMC algorithm shows evident differences from the true pixels, as shown in Figs. 10(b) and 11(b). However, the algorithms based on deep learning show faithful correction results following the slowly varying gray levels.

Fig. 10.

Example of corrected defects for the $3\times 3$ case and the MSE values per pixel. The center part of $3\times 3\text{pixels}$ is estimated. (a) Image block without pixel defects, (b) corrected by TMC (9,221), (c) corrected by the single-layer ANN (165.2), and (d) corrected by Concatenate CNN (649.6).

Fig. 11.

Example of corrected defects for the $5\times 5\text{pixel}$ case and the MSE values per pixel. The center part of $5\times 5\text{pixels}$ is estimated. (a) Image block without pixel defects, (b) corrected by TMC (9,358), (c) corrected by the single-layer ANN (571.6), and (d) corrected by concatenate CNN (11,910).

4. Conclusion

In this paper, we considered several algorithms based on deep learning to correct the pixel defects that occur on the TFT flat panels of radiography detectors. Here, the maximally considered pixel defect sizes were $3\times 3$ and $5\times 5\text{pixels}$. We showed that pixel-defect correction algorithms based on deep learning outperformed the conventional TMC algorithm. Concatenate CNN showed the best performance among the considered neural network models. The performance of the simplest algorithm, which was based on the single-layer ANN model, showed better performance than both the cases of CNN and GAN. Therefore, the single-layer ANN model achieves simple and good performance. Based on the single-layer ANN, correction algorithms for the line defect can be developed to construct practical pixel-defect corrections.

Biographies

Eunae Lee is a PhD candidate in the Department of Electronics Engineering, Hankuk University of Foreign Studies, Republic of Korea. She received her BS and MS degrees from the Hankuk University of Foreign Studies, Republic of Korea, in 2014 and 2017, respectively. Her research interests include statistical signal processing, medical imaging, and deep learning. She was the co-recipient of the 2019 Autumn Conference of Inst. Electr. Inform. Eng. (IEIE) Best Paper Prize.

Eunyeong Hong is with DRTECH Co. Ltd. She received her BS degree from Kyonggi University, Republic of Korea, in 2016, and MS degree from the Department of Electronics Engineering, Hankuk University of Foreign Studies, Republic of Korea in 2020. Her research interests include statistical signal processing, medical imaging, and deep learning.

Dong Sik Kim is a professor in the Department of Electronics Engineering, Hankuk University of Foreign Studies, Republic of Korea. He received his BS, MS, and PhD degrees from Seoul National University, Republic of Korea, in 1986, 1988, and 1994, respectively, all in electrical engineering. His research interests include the theory of quantization, statistical signal processing, medical physics, sensor networks, and smart grid. He was the co-recipient of the 2003 International Workshop on Digital Watermarking (IWDW) Best Paper Prize and the co-recipient of the 2019 Autumn Conference of Inst. Electr. Inform. Eng. (IEIE) Best Paper Prize.

Disclosures

The authors declare no conflicts of interest, financial or otherwise.