文獻探討_All

###### tags: `陳` # 文獻探討_All ## 陳 ### (4) Exploiting ensemble learning for automatic cataract detection and grading #### February 2016 #### Ji-Jiang Yang... - ALL - 眼底相機、1239圖像(767非) - 灰階、直方圖均衡化 - 支援向量機、反向傳播神經網路 - 疊加、多數投票組合出學習模型 - On the other hand, common methods for cataract diagnosis require a slit lamp (e.g., the lens opacity classification system (LOCS III) [4] for clinical assessment or the Wisconsin cataract grading system (Wisconsin System) [5] for photographic grading), which is complicated and expensive for many patients. - [4] The lens opacities classification system III - [5] Cataract prevalence varies substantially with assessment systems: comparison of clinical and photographic grading in a population-based study - 常見的白內障診斷方法需要裂隙燈或其他白內障分級系統，對許多患者來說複雜且昂貴。 - Therefore, reducing costs and simplifying the process for early cataract diagnosis is a crucial means of improving eye care service in less developed areas and bringing light to cataract patients. - 降低成本並簡化早期診斷過程，對低度發展地區的患者來說非常關鍵，並且為他們帶來光明 - We know that ophthalmologists can diagnose cataracts by checking the clear degree of fundus images [3] - [3] Exploiting ensemble learning for automatic cataract detection and grading - 眼科醫師可以透過檢查眼底相片的清晰程度來診斷白內障 - Thus, the task of fundus image classification is to build a classifier to simulate fundus image checking activities for automatic cataract detection and grading [1] - [1] Retinal imaging and image analysis - 眼底圖像分類的目標是建立一個分類器來模擬眼底圖像，檢查(自動)白內障檢測和分級的過程 - Through a combination of three feature extraction methods (i.e., sketch, wavelet, and texture features) and two base learning models (i.e., SVM and BPNN), a total of six independent base classifiers are built. - 眼底圖像分類的目標是建立一個分類器來模擬眼底圖像，檢查(自動)白內障檢測和分級的過程 - Unlike existing work on fundus image classification using a single learning model for cataract detection and grading, this paper proposes an ensemble learning approach that combines multiple heterogeneous learning models for more accurate prediction of cataract classification. - /不同於現有「使用單一學習模型進行白內障檢測和分級的眼底圖像分類」研究，本文提出了一種集成學習方法，將多個相異模型結合，更準確地預測及分類白內障 - 提出一種結合多種學習模型的集合學習方法 - Then, stacking and majority voting are adopted to combine the multiple base classifiers for improved fundus image classification. - /採用疊加和多數投票的方法將多個基本分類器組合，來改善眼底圖像分類。 - 疊加、多數投票組合 ### (5) Mobile Application Based Cataract Detection System #### 23-25 April 2019 #### Vaibhav Agarwal... - ALL - OperCV - KNN分類 - KNN better than SVM & NaiveBayes(不一定用得上) - Geert et.al [11] in their paper entitled “A survey on deep learning in medical image analysis” have surveyed about deep learning algorithms. Deep learning algorithms are convolutional networks have significantly become a method of analyzing and studying medical images... - [11] A survey on deep learning in medical image analysis - 卷積網絡這樣的深度學習演算法，顯然成為了分析和研究醫學圖像的方法 - The use and benefit of a deep learning algorithm are image classification, segmentation, object detection, registration, etc. - [11] A survey on deep learning in medical image analysis - 深度學習演算法的用途和好處在於圖像分類、分割、物件偵測及定位等 - A. Criminisi [14] in their paper entitled “ Machine Learning for Medical Image Analysis” has discussed the application and usage of machine learning in the analysis of images of the medical field - [14] Machine Learning for Medical Image Analysis - 機器學習在醫療上的應用 - The proposed methodology is compared with SVM and Naïve Bayes, as the result shows the KNN based methodology outperforms - KNN > ... - CNN > KNN ? - 將提出的方法與SVM和單純背式分類器進行比較，結果顯示基於KNN的方法的性能表現較為出色 --- ## 王 ### (3) A Robust Automated Cataract Detection Algorithm Using Diagnostic Opinion Based Parameter Thresholding for Telemedicine Application #### Published: 15 September 2016 - Currently, methods available for cataract detection are based on the use of either fundus camera or Digital Single-Lens Reflex (DSLR) camera; both are very expensive. - 現在可以檢測白內障的設備都很昂貴 - /現今可以用於檢測白內障的設備是基於眼底照相機或數位單鏡反射照相機，兩者都很昂貴 - A common health monitoring web platform interfaced with suitable imaging modules will help doctors from remote locations to examine the patient and give a final decision.This way, a more reliable diagnosis can be made. - health monitoring web platform(平台)可協助偏鄉醫師做出判斷 - 健康監測網站平台結合適當的圖像模組後，將幫助偏遠地區的醫生檢查患者並做出最終決定，從而做出更可靠的診斷 - The World Health Report regarding cataracts, updated in 2014, says that 285 million people are estimated to be visually impaired worldwide, out of those 39 million are blind and 246 have low vision.[2] - [2] Visual Impairment and Blindness. Available online: http://www.who.int/mediacentre/factsheets/fs282/en/ (accessed on 20 April 2016). - /2014年更新的《關於白內障的世界衛生報告》聲稱全世界估計有2.8 億視力障礙者，其中3900萬人是盲人，246人視力低下 - In the current scenario, cases of cataracts leading to blindness are likely to advance due to an ageing population and shortage of required healthcare infrastructure in low and middle-income countries. - [3]Gary, B.; Taylor, H. Cataract Blindness–Challenges for 21st century. Bull. World Health Organ. 2001. Available online: http://www.who.int/bulletin/archives/79(3)249.pdf (accessed on 30 April 2016). - 因為白內障失明的人口增加，在中收入國家地所需的醫療設施不足，會使醫療費用增長 - /依照目前的情況，由於人口老齡化和中低收入國家所需醫療基礎設施的短缺，導致失明的白內障病例很可能會增加。 ### (3)-10 Automated Classification of Normal, Cataract and Post Cataract Optical Eye Images using SVM Classifier #### 23-25 October, 2013 - In 1989 , Nuclear Magnetic Resonance (NMR) microscopic imaging method is used to detect the early stages of cataract, but this method is limited by the resolution problem[5]. - [5] Proton magnetic resonance imaging of the ocular lens - 早期的檢測方法有所限制 - /1989年核磁共振顯微鏡成像法被用於檢測初期階段的白內障，但該方法受到分辨率問題的限制 - detect cataract - The color at the inner surface of the cornea is not the same in all the three kinds of images. In cataract images the inner surface of the cornea images is more whitish as compared to that of normal and post-cataract images - /全部三種圖像中，角膜內的表面顏色都不相同。對比白內障圖像與正常圖像，角膜的表面是發白的 - In cataract images, the outer surface of the cornea images is bright in color as compared to that of the normal and Post-cataract images. - 白內障和正常眼睛的角膜顏色不同 - /白內障圖像對比正常圖像，角膜圖像的表面顏色較亮 - Environmental conditions such as the reflection of light influences the quality of the optical images and hence, the efficiency of classification. - 環境條件(如光反射)會影響圖像的質量，因此會影響分類的效率。 - /環境條件(如光反射)會影響圖像的品質，影響分類的效率 - The accuracy of these tools depend on several factors such as the size and quality of the training set, the rigor of the training imparted, and parameters chosen to represent the input - 訓練會因各種不同原因影響準確率 - /模型訓練的準確性取決於幾個因素，例如訓練集的大小和質量、嚴謹性及選擇的參數 - SVM分類 (我們應該用不到) --- ## 洪 ### (6) Cataract Detection Using Single Layer Perceptron Based on Smartphone #### 29-30 Oct. 2019 - This chapter discusses the method used to create a cataract detection system using a single perceptron on a smartphone. - 步驟: 1. Input Image 2. Preprocessing(Cropping Image、Grayscale、 Median Filter) 3. Segmentation(Canny Edge Detection、Hough Circle Transform) 4. Feature Extraction(Mean Intensity、Uniformity) 5. Classification(kNN) - 相關工作 - The preprocessing process starts with the process of cropping and resizing the image, do a Grayscale, and the Median filter. - /預處理從裁剪和調整圖像大小，執行灰度和中值過濾的過程開始 - The segmentation process uses several methods, namely Canny Edge detection, Hough circular transform, and cropping on the pupil area. - /分割過程使用了幾種方法，Canny邊緣檢測、霍夫轉換和瞳孔區域裁剪 - For the feature extraction stage using the gray histogram and get features in the form of mean intensity and uniformity. From the extraction results, the training data uses the single layer perceptron method, then used to determine the classification results in several eye forms, such as normal eyes, immature cataract eyes, and mature cataract eyes. - /使用灰色長條圖的特徵提取階段，以平均強度和均勻性獲得特徵。從提取結果訓練數據使用單層感知機方法，用於確定幾種分類結果，例如:正常眼睛、未成熟白內障眼和成熟白內障 - This system can be used for cataract examination by everyone. If it is detected cataracts are expected to immediately conduct further examinations to the ophthalmologist to get treatment before getting severe. The difference between this system and previous research is that this system uses more training data and data is taken directly from cataract patients with smartphone. - /這個系統可以用於所有人的白內障檢查。如果檢測到白內障，可以在變嚴重前，立即向眼科醫生尋求進一步檢查並治療。這個系統與先前研究之間的區別在於:使用了更多的訓練數據，且數據直接來自擁有智慧型手機的患者 - Some parameters that affect the results of accuracy using KNN as a classification method, namely the relevant features to be used and the k value of K-NN. - KNN會影響某些參數 - /使用KNN分類會影響結果的一些參數，即要使用之相關特徵和K-NN的k值 ### (7)Machine Learning on Cataracts Classification Using SqueezeNet ### 21-24 Oct. 2018 - To classify different areas of cataracts in lens, we use supervised training of convolutional neural network to train 420 images of cataracts on the lens taken from slit-lamps. - 使用CNN訓練 - /為了對水晶體中白內障的不同區域進行分類，我們使用卷積神經網絡的監督式學習訓練:透過裂隙燈拍攝的水晶體上的420張白內障圖像 - The experiment can make the future of classifying cataracts more easily and ophthalmologists can apply operations to different categories of cataracts within a shorter time to cure patients with cataracts. For those people in the countryside, even not so experienced doctors can take the photo of lens and use the program to classify cataracts correctly. - 該實驗使白內障分類更加輕鬆、眼科醫生可以在更短的時間內對不同類別的白內障進行手術，以治愈白內障患者 - <補>對於在偏遠地區的人來說，即使是經驗較少的醫生，也可以使用照相機，並利用程式正確地對白內障進行分類 - First we got our dataset with labels for training and did some image preprocessing such as cropping and image generating methods. - 對數據集圖像做預處理 EX.Cropping、mage generating method - We then trained new neural network layer on top of SqueezeNet, which is a pre-trained model that already achieved high accuracy on classifying images in ImageNet. - 在SqueezeNet上訓練新的神經網路，SqueezeNet是一種預先訓練的模型，已經在ImageNet中對圖像進行分類時實現了高準確度 - SqueezeNet is one type of convolutional neural network (CNN), which achieves the same accuracy on classifying images from ImageNet as AlexNet but needs 50 times fewer parameters than AlexNet. It can be compressed to 510 times smaller than AlexNet. This provided a quicker and more accurate result of the training model. - SqueezeNet是卷積神經網絡（CNN）的一種，它在對ImageNet的圖像進行分類時達到與AlexNet相同的精度，但所需參數比AlexNet少50倍。它可以壓縮到比AlexNet小510倍。這為訓練模型提供了更快，更準確的結果。 - SqueezeNet與其他CNN模型的不同之處在於，有一個構建模塊稱為火災模塊。 --- ## 賴 ### (1) Cataract Detection using Digital Image Processing #### Oct 18-20, 2019 **Two Method** - First (feature extraction based automated detection algorithm) - Preprocessing (used for the extraction of important features) - the image is **cropped** to extract the pupil and is **resized** to enlarge it - **converted to grayscale** and enhanced by applying the Gaussian filter - Feature Extraction - evaluating the texture feature parameter using image **intensity**.The texture feature used for the analysis is **mean(m)**. - **unhealthy eyes will have a higher mean** value of intensity. - Decision Making - based on certain **threshold** values - the mean intensity of a **healthy eye image falls below 50** whereas the ones severely clouded and affected by **cataract have a mean value over 100**. - /健康眼睛圖像的平均強度低於50，嚴重混濁並受白內障影響的圖像的平均值超過100 - Second (Area Calculation) - 前處理的部分都跟第一個方法一樣 - Degree = (cataract area / (pupil area + cataract area)) * 100 - 這方法有講到計算發病的範圍（白內障覆蓋程度），可以延伸我們之前說想做**判讀白內障嚴重程度**。 ## (1-2)Analysis and Study of Cataract Detection Techniques ### 2016 - Cataract is classified into three types based on **the area of protein deposition** named as **nuclear cataract**, **cortical cataract**, **posterior subcapsular cataract**. - /白內障根據蛋白質沉積面積（稱為白內障）、皮層白內障，後囊內白內障分為三類 - Nuclear cataract as the name suggests develops on the **nucleus of the lens**. It is normally caused in **aged people** - /核白內障在晶狀體核上發展，通常是由“老年人”引起的 - Cortical cataract develops in the **cortex of the lens which surrounds the lens** - Posterior sub-capsular cataract develops at the **back of the lens** - /皮質性白內障圍繞在水晶體的皮質中發展 - Preprocessing, Feature Extraction, Classifier construction. - ophthalmoscope(眼底鏡) is used to create the red reflex effect. **A ray of light reflects from the back of the eyeball** and through the pupil to create an optical pathway which helps to detect cataract based on the factors that block this pathway. - /眼底鏡用來產生紅色反射效果。一束光線從眼球的後方反射並穿過瞳孔，形成一條光通路，該光通路可根據阻礙該通路的因素檢測白內障 - Based on the analysis and study of different techniques of cataract detection, it was found that **using fundus images and having Adaboost algorithm as a classifier provided the most accurate results** compared to other techniques that were studied upon. - X -> 眼底圖像+Adaboose最準確 - Slit lamp imaging method is mostly used which is expensive and less accurate. It can only detect and grade a specific class of cataract. The fundus image based cataract system is also studied. Its efficient but not portable and consumes time. Hence there is a **need of a system which has an efficient algorithm and also is portable that is used in smart-phones by remote people with minimum amount of detection time**. ==(我覺得這裡可以延伸因為透過專業器材做預測太困難耗時了，所以我們轉向尋找有無可以從外面直接做判斷的，然後我們就找到～bla bla bla～的paper)== - /裂隙燈成像法是最為常用的，但昂貴且準確性較低。只能檢測和分級特定的白內障。另外研究了基於眼底圖像的白內障系統，效率高但不便於攜帶且浪費時間。因此具有高效的算法的系統，具有便攜性、可讓偏遠地區人員用最少的時間在智能手機中檢測 --- ## 鍾 ### (2) Automatic Cataract Detection of Optical Image using Histogram of Gradient #### June-2018 - classifier constructed by back propagation(反向傳播演算法) - trilateral filter - decrease the noise in the image - feature extraction - GLCM(灰度共生矩陣) - texture feature - image-interval database - linear hough remodel(hough trasform) - edge detect - canny edge detection - eyelids boundary detect - support vector machine - separating hyper plane between two classes which maximizes *margin* --- - SEGMENTATION - Image acquisition: Taking a photograph from iris is that the initial stage of iris primarily based recognition system. - PRE-PROCESSING - The aim of the pre-processing is an improvement of image knowledge that suppresses unwanted distortions or enhances some image options necessary for any process. - SEGMENTATION - Segmentation is the computer version, it’s the method of digital image into multiple segments and its set of pixels, additional referred to as super pixels. --- ### (8) A Hybrid Global-Local Representation CNN Model for Automatic Cataract Grading #### FEBRUARY 2020 - We first had an image preprocessing and selected a deep CNN technique, AlexNet, to learn a global feature representation of the fundus image. Then we used deconvolutional network (DN) in each CNN layer - /首先進行了圖像預處理，選擇一種深層CNN技術-AlexNet，學習眼底圖像的全面特徵表現。然後在每個CNN層中使用反捲積網路 - Preprocessing - Privacy Protection for Patients: As obtained from different fundus cameras, the experimental fundus images have different image sizes. We resized all fundus images uniformly to 256*256 pixels, which makes it suitable for CNN. - /隱私保護：不同的眼底照相機獲得的實驗圖像有不同的大小，將所有眼底圖像的大小統一調整為256*256像素，使其適合CNN - Eliminating Uneven Illumination: Due to local uneven illumination and reflection of eyes, the quality of fundus images is impacted, which may hinder the detection and grading of cataract precisely. Therefore, we converted the original fundus images from RGB color space to the green component images to eliminate the uneven illumination - /消除不均勻的照明：由於局部的不均勻照明和眼睛反射，會影響眼底圖像的品質，可能會影響白內障的檢測和分級的準確。因此，將原始眼底圖像從RGB顏色空間轉換為綠色圖像，以消除不均勻照明 - Feature Extraction by Deep CNN - Deep Convolutional Neural Network (DCNN) is a kind of artificial neural network, which is used for automatically learning features from input images in the field of image recognition. - /深度卷積神經網路是一種人工神經網路，用來從輸入圖像中自動學習特徵 - we selected an extension of a classical deep CNN technique, i.e., AlexNet, to extract features from the retinal fundus images in our experiment. - /選擇經典深層CNN技術（AlexNet）的擴展，在實驗中從視網膜眼底圖像中提取特徵 - Quantify the Prerequisite Features - To quantify the prerequisite features for a fundus image, we attempted to: - interpret the function computed by individual neurons/filters, - examine the overall function computed in convolution layers composed of multiple neurons. - Integrate the Global and Local Feature Representation - with complete fundus image dataset (D1) and variant dataset (D2), the CNN extracts feature sets of different levels: - global feature sets - local feature sets. - we found that CNN tends to capture global feature sets which describe the overall structure of the fundus image, such as shape, texture, edge, and position - /發現CNN傾向於捕捉描述眼底圖像整體結構的全局特徵集，例如形狀、紋理、邊緣和位置 - To combine these different levels of features representation, we designed a new hybrid global-local feature representation model, through which DL technique in DCNN is the basic classifier of ensemble learning - /為了結合不同層次的特徵表現，設計了一個新的混合全局-局部特徵表現模型，透過模型，DCNN中的DL技術成為集成學習的基本分類器 - Constructing Variant Dataset (D2): - To build this global-local feature representation model, we first constructed a variant data set (D2). In this data set, we cut each fundus image in the D1 data set into 8 local 950*950 pixel images to highlight the fundus image. - Combination of the Global and Local Feature Representation - The study have shown that ensemble learning makes the error significantly smaller than any single classifier. In the automatic cataract classification task, we combined the local and global feature-based CNN classifiers by majority voting approach to form an integrated model. - /研究顯示，集成學習讓錯誤明顯小於任何單個分類器。在自動白內障分類任務中，透過多數表決方法結合了基於局部和全局特徵的CNN分類器，形成一個集成模型