1. 程式人生 > >關於視網膜新生血管檢測 部分文獻閱讀筆記

關於視網膜新生血管檢測 部分文獻閱讀筆記

 

包括的文章有:

一、Machine Learning Based Automatic Neovascularization Detection on Optic Disc Region. 

二、Detection of neovascularization in retinal images using semi-supervised learning.

三、Proliferative diabetic retinopathy characterization based on the spatial organization of vascular junctions in fundus images 

四、Automatic detection of neovascularization on optic disk region with feature extraction and support vector machine  5

五、Automated Detection of Neovascularization for Proliferative Diabetic Retinopathy Screening  

六、Detection of neovascularisation using K-means clustering through registration of peripapillary OCT and fundus retinal images(使用K-means聚類通過配準周圍OCT和眼底視網膜影象來檢測新血管形成。)

七、Classification and detection of diabetic retinopathy using K-means algorithm(使用k-means演算法對糖尿病視網膜病變做分類和檢測)

一、Machine Learning Based Automatic Neovascularization Detection on Optic Disc Region

作者:Shuang Yu, Di Xiao and Yogesan Kanagasingam

資料:424幅視網膜影象,134幅有新生血管(NVD),290幅無新生血管(non-NVD)。

準確率:95.23%,特異性:96.30%,敏感性:92.90%

摘要:在本文中,介紹了眼底視網膜影象在視盤區域(NVD)中的新生血管形成的自動檢測。 NV是增生性糖尿病視網膜病變(PDR)發病的指標,其特徵在於視網膜中存在新血管。新生血管很脆弱,造成視力下降的風險很高。因此,不能低估NV的準確和及時檢測的重要性。我們提出了一種用於NVD檢測的自動影象處理過程,包括使用多級Gabor濾波的血管分割,血管形態特徵和紋理特徵的特徵提取,以及支援向量機的影象分類。從每個NVD影象提取42個特徵,並且特徵選擇過程進一步將最佳特徵尺寸減小到18。所選特徵在包含134個NVD和290個非NVD影象的424個視網膜影象上進行訓練和測試。我們平均準確率為95.23%,特異度為96.30%,靈敏度為92.90%,AUC值為98.51%,隨機選擇試驗組。

介紹:糖尿病視網膜病變是青年到中年人致盲和視覺損失的主要原因。糖尿病視網膜病變最嚴重的階段是增生性階段,其重要標誌是新生血管的增殖。新生的血管很細,彎彎曲曲的,而且很脆弱,很容易斷,高風險的導致嚴重的視力損失,甚至失明。

NVD 在視盤包括其附近一個視盤直徑範圍的視網膜出現新生血管,稱為視盤新生血管。

NVE 在其他部分的視網膜新生血管稱為視網膜新生血管。

方法:

  1. 視盤探測:區域性相位對稱;
  2. 預處理:光照校正;非區域性均勻濾波獲得主血管;
  3. 血管分割:血管增強-多規模Gabor濾波器[26]在原圖和主血管圖中應用,得到所有血管和主血管的二值影象-兩圖相減-得到候選微血管。
  4. 特徵提取

1)形態學特徵:血管寬度,長度,曲率,加權長度和加權曲率;統計學特徵,寬度,長度,曲率,加權長度和加權曲率的平均值,方差,偏態[30]。NVD:更小的寬度,更短的長度,更大的曲率,導致偏態值的變化。(15維)+血管數+血管分支數=17維*2=34維(一個在全血管圖,一個在新生血管候選圖)(34維)

曲率=血管畫素長度/兩端點歐氏距離;加權長度=血管畫素長度*血管寬度;加權曲率=血管曲率*血管寬度。

2)紋理特徵:新生血管密度更大,更無組織。提取fractal dimension[31]:分形維數,碎形維度:盒計數方法和傅立葉碎形維度方法。盒計數方法用於三幅二值影象(全部血管,粗血管,候選新生血管), 傅立葉碎形維度用於Gabor濾波後的原圖Fig6(b),和出血管圖Fig6(e)。粒度分析:模式譜的平均值,方差,偏態。(8維)

3)特徵選擇:選擇,加權,最後選出最佳效能的特徵子集。

E.支援向量機,k折交叉分類。

F. NVD資料:

MESSIDOR 4張

High-Resolution Fundus (HRF) 3張

DIARETDB0 11張

Kaggle DR database test 47 train 69

(134張)

Non-NVD資料

290張

(壓縮為視盤半徑為140)

結果和討論:

時間:8.74 second per image.

每類特徵:

特徵選擇:

18個特徵最高

In average, it achieves an accuracy of 95.23%, specificity of 96.30%, sensitivity of 92.90% and AUC of 98.51% on the test set, with the best performance of 95.29% accuracy, 95.16% specificity, 95.65% sensitivity and 99.27% AUC evaluated by the value of AUC.

對比:

精度:

時間:

The accuracy is increased from 91.70% to 95.29% by the feature selection procedure。

二、Detection of neovascularization in retinal images using semi-supervised learning

作者:Appan K. Pujitha*, Gamalapati S. Jahnavi* and Jayanthi Sivaswamy

會議:ISBI 2017

預處理:光照對比度標準化

特徵提取:基於血管特徵的Hessian矩陣。紋理特徵。

三、Proliferative diabetic retinopathy characterization based on the spatial organization of vascular junctions in fundus images

Some efforts have been placed towards automatic characterization of the NVs within the optic disc only (NVDs) [4,5]. NVDs are well contrasted against the background, owing to the fact that the optic disc is the brightest retinal structure. (However, the appearance of NVs, regardless of their site, constitute the transition to the proliferative stage, and it is thus important for the algorithms to be able to detect new vessels in all the available retinal sites.)

方法:

連線點中心探測:血管分割-預處理-張量投票(tensor voting)演算法-連線點中心分離

空間分佈測量:

結果:We apply a 10-fold feature selection process on two different feature sets: (a) including 39 features proposed in the literature, and (b) including 88 features, with 49 newly proposed.

四、Automatic detection of neovascularization on optic disk region with feature extraction and support vector machine

作者:Shuang Yu, Di Xiao and Yogesan Kanagasingam

根據AAO (America Academy of Ophthalmology )的推薦,根據疾病進展階段,DR可分為輕度,中度和重度非增殖性DR(NPDR)和進一步增殖性DR(PDR)。

增殖期是最嚴重的狀態。

視盤探測(區域性相位對稱)-血管分割(多維度Gabor 濾波器)-特徵提取(統計學特徵)

這篇文章是第一篇的會議版。

五、Automated Detection of Neovascularization for Proliferative Diabetic Retinopathy Screening

作者:Sohini Roychowdhury, Dara D. Koozekanani, and Keshab K. Parhi

會議:2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)

摘要:視盤,血管分割,特徵提取(基於區域的特徵),分類。

結果:NVD,sensitivity, specificity and accuracy,74%, 98.2%,87.6%, NVE:61%, 97.5%, 92.1%, PDR篩查,86.4% sensitivity and 76% specificity。

介紹:DR是發展中國家中青年致盲的主要原因。DR分為,非增殖性和增殖性。非增殖期的徵兆有微動脈瘤,棉絨斑,硬性滲出和出血。取決於這些特徵的出現,非增殖期DR 可分為輕度,中等,和重度。

本文貢獻:1、分析比較了5個不同的影象濾波方法的重要性。高通濾波,形態學血管加強,梯度濾波,分水嶺變換,Frangi濾波。高通濾波和形態學血管加強對於NVD/NVE檢查比其他好。2、留一交叉驗證法。選出最優特徵子集。

方法:視盤分割(基於區域的分類)——血管分割(高通濾波)——

高通濾波: 

形態學血管加強:

梯度濾波:

分水嶺變換:

Frangi濾波:

資料庫:STARE:30 normal and 10 images with PDR. Local:All 17 images show varying severities of PDR.

While manifestations of NVD are visible as fine blood vessels in and around the major blood vessels in the OD region, instances of NVE appear as fine vessel-like abnormalities away from the OD region.

結果:

 

 

六、Detection of neovascularisation using K-means clustering through registration of peripapillary OCT and fundus retinal images(使用K-means聚類通過配準周圍OCT和眼底視網膜影象來檢測新血管形成。)

作者:N. Padmasini;R. Umamaheswari(印度人)

會議:2016 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC)

摘要:加強OCT圖和眼底圖,通過kirsch模板提取血管。兩個圖形通過基於配準的相似性測量融合。本文提出使用Kmeans聚類提取NV 特徵來探測正常和不正常的血管。結果通過了所有實時樣本,並得到醫生的肯定。

Investigation of the blood vessels on the retinal images may have multiple applications from being pointers of different retinal diseases.

The Kirsch edge detection algorithm uses a single mask of size 3x3 and rotates it in 45 degree increments through all 8 directions.

高帽變換加強血管。

統計學引數:自相關,曲率,moment invariants(不變距),cluster prominence and cluster shade(簇突出和簇蔭)。

分類器:K-means聚類

 

七、Classification and detection of diabetic retinopathy using K-means algorithm(使用k-means演算法對糖尿病視網膜病變做分類和檢測)

作者:S. B. ManojKumar,H. S. Sheshadri

會議:2016 International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT)

本文摘要:本文主要目的是一個識別背景型糖尿病視網膜病變和增值型糖尿病視網膜病變的框架。