TWI716762B - A method of using neural network to remove tooth image noise - Google Patents
A method of using neural network to remove tooth image noise Download PDFInfo
- Publication number
- TWI716762B TWI716762B TW107139778A TW107139778A TWI716762B TW I716762 B TWI716762 B TW I716762B TW 107139778 A TW107139778 A TW 107139778A TW 107139778 A TW107139778 A TW 107139778A TW I716762 B TWI716762 B TW I716762B
- Authority
- TW
- Taiwan
- Prior art keywords
- noise
- image data
- image
- training
- neural network
- Prior art date
Links
Images
Landscapes
- Apparatus For Radiation Diagnosis (AREA)
- Image Analysis (AREA)
Abstract
一種利用神經網路去除牙齒影像雜訊之方法，步驟包括：(A)讀取一牙齒影像，將該牙齒影像經一影像預處理後獲得一輸入影像資訊；(B)將該輸入影像資訊輸入一已學習之卷積神經網路；(C)該已學習之卷積神經網路過濾雜訊後獲得一去除雜訊之牙齒影像；其中，該已學習之卷積神經網路包含一雜訊特徵及一有效特徵。藉此，可有效改善低劑量X光影像雜訊的問題，產出高品質的牙齒影像。 A method for removing noise from a tooth image using a neural network. The steps include: (A) reading a tooth image, and preprocessing the tooth image to obtain an input image information; (B) inputting the input image information A learned convolutional neural network; (C) the learned convolutional neural network filters noise and obtains a noise-removed tooth image; wherein, the learned convolutional neural network contains a noise Features and an effective feature. In this way, the problem of low-dose X-ray image noise can be effectively improved, and high-quality dental images can be produced.
Description
本發明係關於一種影像處理方法，特別是關於一種利用一種利用神經網路去除牙齒影像雜訊之影像處理方法。 The present invention relates to an image processing method, in particular to an image processing method that utilizes a neural network to remove tooth image noise.
影像處理的相關研究一直是學界、業界所重視的研究課題，尤其是半導體、面板等缺陷檢測，及生物、生態、農業及醫學等影像的各種應用中，常需要檢測出目標的形態、形狀、邊界等，以確認出物體的大小、範圍，尤其在現今醫學影像處理應用中，常常需要處理三維立體影像，因此，三維立體影像的影像處理更顯重要，但其技術難度也大於二維影像的處理。 Image processing related research has always been a research topic that the academic community and the industry attach importance to. Especially in the detection of defects such as semiconductors and panels, and various applications of biological, ecological, agricultural and medical imaging, it is often necessary to detect the shape, shape, and shape of the target. Boundary, etc., to confirm the size and scope of the object. Especially in modern medical image processing applications, it is often necessary to process three-dimensional images. Therefore, the image processing of three-dimensional images is more important, but its technical difficulty is also greater than that of two-dimensional images. deal with.
實務上，目前牙醫師製作假牙的程序繁複，且易使患者不舒適，若能以數位的方式取得單牙立體影像，預期可大幅縮短診斷、治療及製作假牙的時程，但市面上一般是以錐狀射束電腦斷層(CBCT)機台取得全口牙齒的3D立體影像，該影像取得方式需將X光源繞著頭部進行360度密集的拍攝，輻射劑量高，且重建出來的全口牙齒影像其單顆牙齒的解析度不高，無法滿足牙醫師製作假牙的需求，再者， 錐狀射束電腦斷層機台的造價昂貴，無法普及全台約6500家牙科診所，故現有的牙齒立體影像重建技術，有劑量高、解析度低、成本貴等三項缺點。 In practice, the current procedures for dentists to make dentures are complicated and easily make patients uncomfortable. If a single-tooth stereo image can be obtained digitally, it is expected to greatly shorten the time course of diagnosis, treatment and denture making, but the market is generally Use a cone beam computer tomography (CBCT) machine to obtain 3D stereoscopic images of the full mouth teeth. This image acquisition method requires 360-degree intensive shooting of the X light source around the head, high radiation dose, and the reconstructed full mouth The resolution of the tooth image of a single tooth is not high, which cannot meet the needs of dentists for making dentures. Furthermore, Cone-beam computed tomography machines are expensive to build and cannot be used in about 6,500 dental clinics in Taiwan. Therefore, the existing three-dimensional tooth image reconstruction technology has three shortcomings: high dose, low resolution, and high cost.
低X光劑量的牙齒立體影像因此備受矚目，其係利用X光影像感測器置放於口腔中，與口腔外的X光源進行相對旋轉運動，在小角度的X光掃描範圍內，進行較低密度的連續拍攝，以降低X光的劑量，但低X光劑量的牙齒立體影像卻因為能量較低，相對地雜訊問題就變得會影響牙齒影像品質；而處理雜運的先前技術，中值濾波器為常用的非線性濾波方法，也是圖像處理技術中最常用的預處理技術，其基本原理是選擇待處理像素的一個鄰域中各像素值的中值來代替待處理的像素，其主要功能是像素的灰度值與周圍像素比較接近，從而消除孤立的雜訊點，所以中值濾波器能夠很好的消除椒鹽雜訊，但缺點的部分是，根據中值濾波器原理，如果在濾波窗口內的雜訊點的個數大於整個窗口內像素的個數，則中值濾波就不能很好的過濾掉雜訊，換句話說，可以有效的去除比較強的雜訊訊號，相對來說對於低頻雜訊訊號說影響就比較小了，因此較適用於對椒鹽雜訊的處理。 Stereoscopic images of teeth with low X-ray doses have therefore attracted much attention. The X-ray image sensor is placed in the oral cavity and rotates relative to the X-ray source outside the oral cavity. It performs a small-angle X-ray scanning range. Low-density continuous shooting to reduce the X-ray dose, but the low X-ray dose of the three-dimensional image of the tooth is relatively low in energy, and the noise problem will affect the quality of the tooth image; and the prior art to deal with clutter The median filter is a commonly used nonlinear filtering method, and it is also the most commonly used preprocessing technology in image processing technology. The basic principle is to select the median value of each pixel in a neighborhood of the pixel to be processed instead of the pixel to be processed Pixel, its main function is that the gray value of the pixel is closer to the surrounding pixels, thereby eliminating isolated noise points, so the median filter can eliminate salt and pepper noise very well, but the disadvantage is that according to the median filter The principle is that if the number of noise points in the filter window is greater than the number of pixels in the entire window, the median filter cannot filter out the noise well, in other words, it can effectively remove the stronger noise The signal has relatively little effect on low-frequency noise signals, so it is more suitable for the processing of salt and pepper noise.
因此目前業界極需發展出一種利用神經網路去除牙齒影像雜訊之方法，可有效處理影像雜訊的問題，如此一來，方能同時兼具安全與品質，避免高劑量的X光傷害人體，同時有效過濾雜訊維持影像品質，有效增加低劑量X光 在醫學領域及其他相關領域的應用。 Therefore, there is a great need for the industry to develop a method of using neural networks to remove dental image noise, which can effectively deal with the problem of image noise. In this way, it can have both safety and quality, and avoid high-dose X-ray damage to the human body. , While effectively filtering noise to maintain image quality, effectively increasing low-dose X-rays Applications in the medical field and other related fields.
鑒於上述悉知技術之缺點，本發明之主要目的在於提供一種利用神經網路去除牙齒影像雜訊之方法，整合一牙齒影像、一已學習之卷積神經網路、一雜訊特徵及一有效特徵等，以避免雜訊造成的錯誤影響，造成低劑量X光影像的判讀錯誤。 In view of the shortcomings of the above-mentioned known technology, the main purpose of the present invention is to provide a method for removing noise from tooth images using neural networks, integrating a tooth image, a learned convolutional neural network, a noise feature, and an effective method. Features, etc., to avoid erroneous influence caused by noise and cause errors in interpretation of low-dose X-ray images.
為了達到上述目的，根據本發明所提出之一方案，提供一種利用神經網路去除牙齒影像雜訊之方法，步驟包括：(A)讀取一牙齒影像，將該牙齒影像經一影像預處理後獲得一輸入影像資訊；(B)將該輸入影像資訊輸入一已學習之卷積神經網路；(C)該已學習之卷積神經網路過濾雜訊後獲得一去除雜訊之牙齒影像；其中，該已學習之卷積神經網路包含一雜訊特徵及一有效特徵。 In order to achieve the above objective, according to a solution proposed in the present invention, a method for removing noise from a tooth image using a neural network is provided. The steps include: (A) reading a tooth image, and preprocessing the tooth image Obtain an input image information; (B) input the input image information into a learned convolutional neural network; (C) obtain a noise-removed tooth image after filtering the learned convolutional neural network; Among them, the learned convolutional neural network includes a noise feature and an effective feature.
步驟(B)中該已學習之卷積神經網路可含包含一訓練階段及一使用階段，訓練階段使用沒有雜訊的訓練影像資料，及加有高斯分布的雜訊訓練影像資料來篩選、過濾雜訊，藉由有效特徵及雜訊特徵的萃取，可訓練卷積神經網路進行訓練，並結合深度殘差學習概念，使得卷積神經網路可在使用階段有效過濾掉雜訊。 The learned convolutional neural network in step (B) may include a training phase and a use phase. The training phase uses training image data without noise and noise training image data with Gaussian distribution to filter, Filtering noise, through the extraction of effective features and noise features, convolutional neural networks can be trained for training, combined with the concept of deep residual learning, so that convolutional neural networks can effectively filter out noise in the use phase.
上述的的訓練階段，包含一特徵萃取處理，本發明藉由結合深度殘差學習的卷積神經網路的訓練，萃取出有 效特徵及雜訊特徵，可將具有雜訊特徵的雜訊過濾；其中有效特徵萃取的步驟包含：(a)將一訓練影像資料經過一加權後映射得到一加權影像資料、(b)該加權影像資料反向加權映射後得到一訓練完成影像資料、(c)該訓練影像資料與訓練完成影像資料進行誤差疊代以獲得該有效特徵，而該訓練影像資料是一低劑量無雜訊的X光影像資料；而雜訊特徵萃取的步驟包含：(1)將一雜訊訓練影像資料經過一加權後映射得到一加權影像資料、(2)該雜訊加權影像資料反向加權映射後得到一訓練完成影像資料、(3)該訓練影像資料與訓練完成影像資料進行誤差疊代以獲得該雜訊特徵，其中雜訊訓練影像資料是將一高斯分布的雜訊資訊加入上述的訓練影像資料。 The above-mentioned training stage includes a feature extraction process. The present invention uses a convolutional neural network training combined with deep residual learning to extract The effective feature and noise feature can filter the noise with the noise feature; the step of extracting the effective feature includes: (a) mapping a training image data after a weighting to obtain a weighted image data, (b) the weighting After the image data is inversely weighted and mapped, a training completed image data is obtained, (c) the training image data and the training completed image data are error iterated to obtain the effective feature, and the training image data is a low-dose, no-noise X Optical image data; and the steps of extracting noise features include: (1) a weighted image data is mapped to a noise training image data to obtain a weighted image data; (2) the noise weighted image data is reversely weighted to obtain a weighted image data The training completed image data, (3) the training image data and the training completed image data are error iterated to obtain the noise characteristics, wherein the noise training image data is a Gaussian distributed noise information added to the above training image data.
本案發明的牙齒影像可以是一低劑量X光影像，低劑量X光影像可先經一影像預處理，此影像預處理可包含增強對比度、邊緣檢測、二值化處理等程序後，再進行正規化，使低劑量X光影像的每一像素值為落在0到1之間的機率分布，接下來就可進行人工的資料集擴增，包含旋轉、平移、平面翻轉等，以獲取大量的資料(訓練影像資料)來提供模型訓練。 The tooth image of the present invention can be a low-dose X-ray image. The low-dose X-ray image can be pre-processed. This image pre-processing can include contrast enhancement, edge detection, binarization and other procedures, and then regular In this way, the probability distribution of each pixel value of low-dose X-ray images falling between 0 and 1. Then, manual data set amplification can be carried out, including rotation, translation, plane flip, etc., to obtain a large number of Data (training image data) to provide model training.
以上之概述與接下來的詳細說明及附圖，皆是為了能進一步說明本創作達到預定目的所採取的方式、手段及功效。而有關本創作的其他目的及優點，將在後續的說明及圖式中加以闡述。 The above summary and the following detailed description and drawings are for the purpose of further explaining the methods, means and effects of this creation to achieve the intended purpose. The other purposes and advantages of this creation will be explained in the following description and diagrams.
S101-S103‧‧‧步驟 S101-S103‧‧‧Step
x‧‧‧訓練影像資料 x‧‧‧Training image data
y‧‧‧加權影像資料 y‧‧‧weighted image data
z‧‧‧訓練完成影像資料 z‧‧‧Image data of training completed
x’‧‧‧雜訊訓練影像資料 x’‧‧‧Noise training image data
第一圖係為本發明一種利用神經網路去除牙齒影像雜訊之方法流程圖；第二圖係為本發明一種卷積神經網路架構示意圖；第三圖係為本發明一種深度學習模型架構示意圖；第四圖係為本發明一種利用卷積自動編碼器取得有效特徵之流程示意圖；第五圖係為本發明一種利用卷積自動編碼器取得雜訊特徵之流程示意圖；第六圖係為本發明實施例去除雜訊前之低劑量X光影像(a)與去除雜訊後之低劑量X光影像圖(b)。 The first figure is a flow chart of a method for removing noise from a tooth image using a neural network; the second figure is a schematic diagram of a convolutional neural network architecture of the present invention; the third figure is a deep learning model architecture of the present invention Schematic diagram; the fourth diagram is a schematic diagram of the process of obtaining effective features using a convolutional autoencoder of the present invention; the fifth diagram is a schematic diagram of the process of obtaining noise features using a convolutional autoencoder of the present invention; the sixth diagram is The low-dose X-ray image before noise removal (a) and the low-dose X-ray image after noise removal (b) in the embodiment of the present invention.
以下係藉由特定的具體實例說明本創作之實施方式，熟悉此技藝之人士可由本說明書所揭示之內容輕易地了解本創作之優點及功效。 The following is a specific example to illustrate the implementation of this creation. Those who are familiar with this technique can easily understand the advantages and effects of this creation from the content disclosed in this manual.
本發明藉由深度學習應用在影像去雜訊初步模型訓練，所利用的模型架構則是卷積神經網路(convolution neural network,CNN)的變形，分別是卷積自動編碼器(convolution autoencoder,CAE)以及深度殘差網路 (deep residual network,ResNet)，兩者經由訓練之後皆有初步的去雜訊效果，並利用實體牙齒當作模型的學習目標，其中以深度殘差網路有良好的效果，其輸出影像較能保留原來圖像之邊緣完整度，訓練完成後再以低劑量的X光影像當作輸入端，透過訓練好的神經網路來進行去雜訊，可得到高品質的影像。 The present invention uses deep learning to apply to image denoising preliminary model training. The model architecture used is a convolution neural network (convolution neural network, CNN) variant, which is a convolution autoencoder (CAE). ) And deep residual network (deep residual network, ResNet), both have a preliminary denoising effect after training, and use solid teeth as the learning target of the model. Among them, the deep residual network has good results, and the output image is better The edge integrity of the original image is preserved. After the training is completed, the low-dose X-ray image is used as the input, and the trained neural network is used to remove noise, and high-quality images can be obtained.
卷積神經網絡(convolutional neural networks；CNN)是近年發展起來，並引起廣泛重視的一種高效識別方法。1960年代，Hubel和Wiesel在研究貓腦皮層中用於局部敏感和方向選擇的神經元時發現其獨特的網絡結構可以有效地降低反饋神經網絡的複雜性，繼而提出了卷積神經網絡。在屬資料量龐大的醫學影像中，CNN在模式分類領域因為避免了對圖像的複雜前期預處理，可以直接輸入原始圖像，因而在影像上面得到了更為廣泛的應用，在這樣的條件下，本發明以卷積神經網路的架構做為出發，結合結合深度殘差學習概念，改良其他模型架構而得到更好的學習模型；苯發明在原始影像上加入不同標準差分布的人工高斯雜訊做為訓練資料，並以較為清晰、雜訊較少且影像品質較高的影像當作學習目標，透過定義好的神經網路架構來進行模型訓練，使得神經網路經過反覆的誤差疊代後能學習到影像前後的差異性以及輸入資料的關鍵特徵，也就是經過加上雜訊後的這個破損數據，因而達到去雜訊的效用，在資料量不完整 且影像品質較低的情況下，提升影像品質。 Convolutional neural networks (CNN) is an efficient recognition method that has been developed in recent years and has attracted widespread attention. In the 1960s, Hubel and Wiesel discovered that their unique network structure can effectively reduce the complexity of feedback neural networks when studying neurons used for local sensitivity and direction selection in the cat brain cortex, and then proposed a convolutional neural network. In medical images with a huge amount of data, CNN in the field of pattern classification avoids the complicated pre-processing of the image and can directly input the original image. Therefore, it has been more widely used in the image. Under such conditions Next, the present invention takes the architecture of convolutional neural network as the starting point, combines the concept of deep residual learning, and improves other model architectures to obtain a better learning model; Benzene invention adds artificial Gaussian distributions with different standard deviations to the original image Noise is used as training data, and images with clearer, less noise and higher image quality are used as learning targets. Model training is performed through a well-defined neural network architecture, so that the neural network undergoes repeated errors. After the generation, you can learn the difference between before and after the image and the key characteristics of the input data, that is, the damaged data after adding noise, so as to achieve the effect of noise removal, when the amount of data is incomplete And when the image quality is low, the image quality is improved.
請參閱第一圖，為本發明一種利用神經網路去除牙齒影像雜訊之方法流程圖。如圖所示，本發明所提供一種利用神經網路去除牙齒影像雜訊之方法，步驟包括：(A)讀取一牙齒影像，將該牙齒影像經一影像預處理後獲得一輸入影像資訊S101；(B)將該輸入影像資訊輸入一已學習之卷積神經網路S102；(C)該已學習之卷積神經網路過濾雜訊後獲得一去除雜訊之牙齒影像S103；其中，該已學習之卷積神經網路包含一雜訊特徵及一有效特徵。 Please refer to the first figure, which is a flow chart of a method for removing noise from a tooth image using a neural network according to the present invention. As shown in the figure, the present invention provides a method for removing noise from a tooth image by using a neural network. The steps include: (A) reading a tooth image, and preprocessing the tooth image to obtain an input image information S101 (B) the input image information is input into a learned convolutional neural network S102; (C) the learned convolutional neural network filters noise to obtain a noise-removed tooth image S103; wherein, the The learned convolutional neural network includes a noise feature and an effective feature.
實施例 Example
本實施例包含一訓練階段及一使用階段，訓練階段首先第一步先選取出無雜訊的高品質影像資料(低劑量高品質X光牙齒影像)，將這些無雜訊的高品質影像資料集進行正規化，使每一影像資料像素值為落在0到1之間的機率分布，再來進行正規化處理後的資料集進行擴增，擴增的手法包含旋轉、平移、平面翻轉等影像處理方法，以獲取大量的影像資料來提供一神經網路做訓練；請參閱第二圖，為本發明一種卷積神經網路架構示意圖，如圖所示，本發明使用深度殘差學習的概念結合自動編碼器的卷積神經網路架構，該卷積神經網路係為一種有卷積自動編碼器的架構，包含有多層神經網路，其中輸入層和輸出層表示相同的含義，具有相同的節點數，即輸出層的神經元數量完全等於輸入層神經元的數量，隱藏 層的神經元數量少於輸出層的神經元數量，本實施例是將神經網路NN的隱含層看成是一個編碼器和解碼器，輸入資料經過隱含層的編碼和解碼，到達輸出層時，確保輸出的結果儘量與輸入資料保持一致，本實施例是利用卷積層(Convolutional Layer)來取代習知原始自動編碼器，利用Encoder及Decoder的全連接層來有效達到更好的特徵學習；請參閱第三圖，為本發明一種深度學習模型架構示意圖，如圖所示，本發明的state-of-art的卷積神經網路(CNN)架構相當多層，如果只是簡單的堆疊層來因應網路深度，不僅會造成梯度消失的問題，深層的網路也會很難訓練，因為誤差可能過大，本實施例在神經網路中層與層之間引入一個identity shortcut的概念，直接跳過一層與多層，藉由透過這樣一個殘差映射的方式，較深的模型所產生的訓練誤差和較淺的模型務差相近，其訓練結果意謂著較能保留原來影像的特徵及形狀。 This embodiment includes a training phase and a use phase. In the training phase, the first step is to select noise-free high-quality image data (low-dose high-quality X-ray dental images), and combine these noise-free high-quality image data. The set is normalized, so that the probability distribution of each image data pixel value falls between 0 and 1, and then the normalized data set is amplified. The amplification methods include rotation, translation, plane flip, etc. Image processing method to obtain a large amount of image data to provide a neural network for training; please refer to the second figure, which is a schematic diagram of a convolutional neural network architecture of the present invention. As shown in the figure, the present invention uses deep residual learning The concept is combined with the convolutional neural network architecture of the autoencoder. The convolutional neural network is an architecture with a convolutional autoencoder. It includes a multi-layer neural network. The input layer and the output layer have the same meaning. The same number of nodes, that is, the number of neurons in the output layer is exactly equal to the number of neurons in the input layer, hidden The number of neurons in the layer is less than the number of neurons in the output layer. In this embodiment, the hidden layer of the neural network NN is regarded as an encoder and decoder. The input data is encoded and decoded by the hidden layer to reach the output When layering, ensure that the output result is as consistent as possible with the input data. In this embodiment, a convolutional layer is used to replace the conventional original autoencoder, and the fully connected layer of Encoder and Decoder is used to effectively achieve better feature learning. Please refer to the third figure, which is a schematic diagram of the architecture of a deep learning model of the present invention. As shown in the figure, the state-of-art convolutional neural network (CNN) architecture of the present invention is quite multi-layered. If it is just a simple stack of layers In response to the depth of the network, not only will the gradient disappear, but the deep network will also be difficult to train because the error may be too large. This embodiment introduces the concept of identity shortcut between the layers in the neural network, and skips directly One layer and multiple layers, through such a residual mapping method, the training error generated by the deeper model is similar to the training error of the shallower model, and the training result means that the characteristics and shape of the original image can be retained.
請參閱第四圖，為本發明一種利用卷積自動編碼器取得有效特徵之流程示意圖、請參閱第五圖，為本發明一種利用卷積自動編碼器取得雜訊特徵之流程示意圖。如圖所示，本發明實施例利用卷積自動編碼器取得有效特徵之步驟包含：第一步：將一訓練影像資料(x)經過加權(W,b)，映射(RELU)之後得到y(加權影像資料)，此為一編碼程序 (Encoder：y=s(Wx+b))，其中，訓練影像資料(x)係為無雜訊的高品質影像資料(低劑量高品質X光牙齒影像)；第二步：再對y(加權影像資料)反向加權映射後得到一z(訓練完成影像資料)，此為一編碼程序(Decoder：z=s(W'y+b'))；第三步：通過反覆誤差疊代訓練兩組(W,b)，使得誤差函數最小並讓z近似於x而獲得有效特徵。 Please refer to Figure 4, which is a schematic diagram of a process of obtaining effective features using a convolutional autoencoder of the present invention. Please refer to Figure 5, which is a schematic diagram of a process of obtaining noise features using a convolutional autoencoder of the present invention. As shown in the figure, the steps of using a convolutional autoencoder to obtain effective features in the embodiment of the present invention include: The first step: weighting (W, b) a training image data (x), mapping (RELU) to obtain y( Weighted image data), this is an encoding process ( Encoder : y = s ( Wx + b )), where the training image data (x) is noise-free high-quality image data (low-dose high-quality X-ray dental image ); Step 2: Reverse weighted mapping of y (weighted image data) to obtain a z (trained image data), which is an encoding procedure ( Decoder : z = s ( W'y + b' )); The third step: Train two sets of (W, b) through iterative error iteration to minimize the error function and allow z to approximate x to obtain effective features.
同樣地，本發明實施例係利用卷積自動編碼器取得雜訊特徵，其步驟包含：第一步：將一雜訊訓練影像資料(x’)經過一加權後映射得到一y(加權影像資料)，其中，雜訊訓練影像資料(x’)係以隨機機率分布的方式給訓練影像資料(x)的矩陣中隨機加入雜訊值(雜訊值為高斯分布)；第二步：再對y(加權影像資料)反向加權映射後得到一z(訓練完成影像資料)；第三步：通過反覆誤差疊代訓練兩組(W,b)，使得誤差函數最小並讓z近似於x(訓練影像資料)而獲得雜訊特徵。 Similarly, the embodiment of the present invention uses a convolutional autoencoder to obtain noise characteristics. The steps include: The first step is to map a noise training image data (x') after a weighting to obtain a y(weighted image data) ), where the noise training image data (x') is randomly added to the matrix of the training image data (x) in a random probability distribution (the noise value is Gaussian distribution); the second step: correct again y (weighted image data) reverse weighted mapping to obtain a z (training completed image data); the third step: train two sets of (W, b) through iterative error iteration to minimize the error function and make z approximate to x( Training image data) to obtain noise characteristics.
本實施例經過上述訓練階段獲得一已學習之卷積神經網路後，可進入使用階段，首先讀取一牙齒影像，此牙齒影像係為一低劑量X光影像，將該牙齒影像經一影像預處理(如增強對比度、邊緣檢測、二值化、正規化等處理)後，獲得一可應用於該已學習之卷積神經網路的輸入影像資訊，該已學習之卷積神經網路利用雜訊特徵及有效特徵，即 可有效過濾掉雜訊而獲得一去除雜訊之牙齒影像。 After obtaining a learned convolutional neural network through the above-mentioned training phase, this embodiment can enter the use phase. First, read a tooth image. The tooth image is a low-dose X-ray image, and the tooth image is passed through an image After preprocessing (such as contrast enhancement, edge detection, binarization, normalization, etc.), an input image information that can be applied to the learned convolutional neural network is obtained, and the learned convolutional neural network is used Noise characteristics and effective characteristics, namely It can effectively filter out the noise and obtain a tooth image without noise.
請參閱第六圖，為本發明實施例去除雜訊前之低劑量X光影像(a)與去除雜訊後之低劑量X光影像圖(b)。如圖所示，本發明利用低劑量影像經神經網路去雜訊後的結果，使用門牙作為待測物，比較原重建影像與去雜訊後之影像之對比解析度CNR(contrast-to-noise ratio)與訊雜比SNR(signal-to-noise ratio)，可以發現經去雜訊後的影像(圖b)確實是比前者(圖a)的影像品質來得佳，根據表一的定量分析數據顯示，也能經由2個檢測位置再次比較出兩者的雜訊程度，去雜訊過後的2個檢測位置之CNR及SNR比的確較高，影像品質確實有提升。 Please refer to the sixth figure, which shows the low-dose X-ray image (a) before noise removal and the low-dose X-ray image after noise removal (b) according to the embodiment of the present invention. As shown in the figure, the present invention uses the result of low-dose image denoising by neural network, using incisor teeth as the test object, and comparing the contrast resolution CNR (contrast-to- Noise ratio) and signal-to-noise ratio (SNR), it can be found that the image after denoising (Figure b) is indeed better than the former (Figure a). According to the quantitative analysis in Table 1. The data shows that the noise levels of the two can be compared again through the two detection positions. The CNR and SNR ratios of the two detection positions after the noise removal are indeed higher, and the image quality is indeed improved.
本發明解決在低密度X光拍攝下所產生的假影問題，在影像後處理的部份，採用人工智慧中深度學習的方式，去訓練出一套神經網路，來進行自動去雜訊，再次提昇重建後的影像品質，可提供牙醫師在製作假牙或醫療過程中所需的牙齒資訊。 The present invention solves the problem of artifacts generated under low-density X-ray shooting. In the image post-processing part, the deep learning method in artificial intelligence is used to train a set of neural networks to automatically remove noise. Improve the quality of the reconstructed image again, which can provide dental information needed by the dentist during the production of dentures or medical procedures.
上述之實施例僅為例示性說明本創作之特點及 功效，非用以限制本創作之實質技術內容的範圍。任何熟悉此技藝之人士均可在不違背創作之精神及範疇下，對上述實施例進行修飾與變化，因此，本創作之權利保護範圍，應如後述之申請專利範圍所列。 The above-mentioned embodiments are only illustrative to illustrate the characteristics of this creation and The effect is not to limit the scope of the actual technical content of this creation. Anyone familiar with this technique can modify and change the above-mentioned embodiments without violating the spirit and scope of creation. Therefore, the scope of protection of the rights of this creation should be listed in the scope of patent application described later.
S101-S103‧‧‧步驟 S101-S103‧‧‧Step
Claims (1)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW107139778A TWI716762B (en) | 2018-11-07 | 2018-11-07 | A method of using neural network to remove tooth image noise |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW107139778A TWI716762B (en) | 2018-11-07 | 2018-11-07 | A method of using neural network to remove tooth image noise |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| TW202018658A TW202018658A (en) | 2020-05-16 |
| TWI716762B true TWI716762B (en) | 2021-01-21 |
Family
ID=71895808
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| TW107139778A TWI716762B (en) | 2018-11-07 | 2018-11-07 | A method of using neural network to remove tooth image noise |
Country Status (1)
| Country | Link |
|---|---|
| TW (1) | TWI716762B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111950584B (en) * | 2020-06-16 | 2024-05-14 | 江西中科九峰智慧医疗科技有限公司 | Intelligent identification method and system for position integrity in X-ray chest radiography |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW514513B (en) * | 1996-02-06 | 2002-12-21 | Deus Technologies Inc | Method for the detection of lung nodule in radiological images using digital image processing and artificial neural network |
| TW201117130A (en) * | 2009-11-12 | 2011-05-16 | Univ Nat Kaohsiung Applied Sci | Method and system for cancelling noise in gray-level images and computer-readable medium thereof |
| TWI385598B (en) * | 2009-12-01 | 2013-02-11 | Ind Tech Res Inst | Image processing method, training method of classifier and evaluate method of lesion risk |
| US20160174902A1 (en) * | 2013-10-17 | 2016-06-23 | Siemens Aktiengesellschaft | Method and System for Anatomical Object Detection Using Marginal Space Deep Neural Networks |
| EP3326571A1 (en) * | 2016-11-25 | 2018-05-30 | RWTH Aachen University | Method and system for reconstructing dental structures |
| US20180247193A1 (en) * | 2017-02-24 | 2018-08-30 | Xtract Technologies Inc. | Neural network training using compressed inputs |
| CN108566537A (en) * | 2018-05-16 | 2018-09-21 | 中国科学院计算技术研究所 | Image processing apparatus for carrying out neural network computing to video frame |
| CN108647660A (en) * | 2018-05-16 | 2018-10-12 | 中国科学院计算技术研究所 | A method of handling image using neural network chip |
| WO2018200493A1 (en) * | 2017-04-25 | 2018-11-01 | The Board Of Trustees Of The Leland Stanford Junior University | Dose reduction for medical imaging using deep convolutional neural networks |
-
2018
- 2018-11-07 TW TW107139778A patent/TWI716762B/en active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW514513B (en) * | 1996-02-06 | 2002-12-21 | Deus Technologies Inc | Method for the detection of lung nodule in radiological images using digital image processing and artificial neural network |
| TW201117130A (en) * | 2009-11-12 | 2011-05-16 | Univ Nat Kaohsiung Applied Sci | Method and system for cancelling noise in gray-level images and computer-readable medium thereof |
| TWI385598B (en) * | 2009-12-01 | 2013-02-11 | Ind Tech Res Inst | Image processing method, training method of classifier and evaluate method of lesion risk |
| US20160174902A1 (en) * | 2013-10-17 | 2016-06-23 | Siemens Aktiengesellschaft | Method and System for Anatomical Object Detection Using Marginal Space Deep Neural Networks |
| EP3326571A1 (en) * | 2016-11-25 | 2018-05-30 | RWTH Aachen University | Method and system for reconstructing dental structures |
| US20180247193A1 (en) * | 2017-02-24 | 2018-08-30 | Xtract Technologies Inc. | Neural network training using compressed inputs |
| WO2018200493A1 (en) * | 2017-04-25 | 2018-11-01 | The Board Of Trustees Of The Leland Stanford Junior University | Dose reduction for medical imaging using deep convolutional neural networks |
| CN108566537A (en) * | 2018-05-16 | 2018-09-21 | 中国科学院计算技术研究所 | Image processing apparatus for carrying out neural network computing to video frame |
| CN108647660A (en) * | 2018-05-16 | 2018-10-12 | 中国科学院计算技术研究所 | A method of handling image using neural network chip |
Also Published As
| Publication number | Publication date |
|---|---|
| TW202018658A (en) | 2020-05-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11727569B2 (en) | Training a CNN with pseudo ground truth for CT artifact reduction | |
| Hatvani et al. | Deep learning-based super-resolution applied to dental computed tomography | |
| CN111325686B (en) | Low-dose PET three-dimensional reconstruction method based on deep learning | |
| JP2020168352A (en) | Medical apparatus and program | |
| CN113487657B (en) | Deep learning-based mode conversion method | |
| CN110728729B (en) | Attention mechanism-based unsupervised CT projection domain data recovery method | |
| CN111047524A (en) | Low-dose CT lung image denoising method based on deep convolutional neural network | |
| Kumar et al. | Descriptive analysis of dental X-ray images using various practical methods: A review | |
| CN112822982B (en) | Image forming apparatus, image forming method, and method for forming learning model | |
| CN110675461A (en) | CT image recovery method based on unsupervised learning | |
| WO2022110530A1 (en) | Tomographic image reconstruction method based on spect data sampling and noise characteristics | |
| US20230196515A1 (en) | A method of denoising dental images through domain adaptation | |
| CN110473150B (en) | CNN medical CT image denoising method based on multi-feature extraction | |
| US9672641B2 (en) | Method, apparatus, and computer readable medium for removing unwanted objects from a tomogram | |
| Zhu et al. | Metal artifact reduction for X-ray computed tomography using U-net in image domain | |
| CN113706409A (en) | CBCT image enhancement method and device based on artificial intelligence and storage medium | |
| CN104809713B (en) | The non-linear sharpening enhancement method of CBCT panorama sketch based on neighborhood information and gaussian filtering | |
| CN112489158A (en) | Enhancement method for low-dose PET image by using cGAN-based adaptive network | |
| CN110101401B (en) | Liver contrast agent digital subtraction angiography method | |
| CN116645283A (en) | Low-dose CT image denoising method based on self-supervision perceptual loss multi-scale convolutional neural network | |
| TWI716762B (en) | A method of using neural network to remove tooth image noise | |
| Wang et al. | Texture‐aware dual domain mapping model for low‐dose CT reconstruction | |
| CN112258438B (en) | LDCT image recovery method based on unpaired data | |
| Johari et al. | Metal artifact suppression in dental cone beam computed tomography images using image processing techniques | |
| CN110176045A (en) | A method of dual-energy CT image is generated by single energy CT image |