醫療影像處理在診斷上之應用 嘉義大學資工系 教授 柯建全 時間 : 2009 年 11 月 13 日.

Presentation on theme: "醫療影像處理在診斷上之應用 嘉義大學資工系 教授 柯建全 時間 : 2009 年 11 月 13 日."— Presentation transcript:

1 醫療影像處理在診斷上之應用 嘉義大學資工系 教授 柯建全 時間 : 2009 年 11 月 13 日

2 Outline  Introduction  Object of medical image processing  Imaging devices  applications  Related techniques for Medical imaging  Research Results  Future works

3 Introduction  What is Medical imaging?  Why do we need digital image processing?  What kind of problems are often caused in medical images? Blurring caused by respiratory or motion Low contrast caused by imaging device or resolution Complicated textures  Research trends have been transferred from 2-D to 3-D reconstruction

4 Introduction (continue)  Integrate all possible methods in the filed of DIP, pattern recognition, and computer graphics  Qualitative  Quantitative  Three categories of imaging in different modalities Structural image Functional image Molecular image

5 Object  Help physicians diagnose Reduce inter- and intra-variability  Produce qualitative and quantitative assessment by computer technologies  Determine appropriate treatments according to the analyses  Surgical simulation or skills to reduce possible errors

6 Medical Imaging Modalities  X-ray  Ultrasound: non-invasive  Computed tomography  Magnetic resonance imaging  SPECT (Single photon emission tomography)  PET( Positron emission tomography)  Microscopy

7

8 X-ray

9 Ultrasound  2-D sonography  3-D sonography  Doppler color sonography A series of 2-D projection Reconstruction  4-D sonography

10

11

12 Computed tomography

13 MRI  可以觀察活體三度空間的斷層影像  磁振影像取影像時可以適當控制而得到不 同參數的影像，如溫度、流場 (flow) 、水 含量、分子擴散 ( diffusion) 、 灌流 (perfusion) 、化學位移 (chemical shift) 、 功能性 (functional MRI) 及不同核種如 氫、碳、磷

14 MRI-structural and functional image

15

16 Related techniques  Image processing Segmentation Registration Feature Extraction  Shape feature  Texture Motion tracking  Pattern recognition Supervised learning Un-supervised learning Neuro network Fuzzy Support vector machine(SVM) Genetic algorithm

17 Related techniques  3-D graphic Virtual diagnose or visualization Fusion between different modalities Bio-medical visualization

18 SPECT-functional image

19 PET(Positron Emission Tomography )  PET 以分子細胞學為基礎，將帶有特殊標記的葡 萄糖合成藥劑注入受檢者體內，利用 PET 掃瞄儀 的高解析度與靈敏度作全身的掃描，藉由癌細胞 分裂迅速，新陳代謝特別旺盛，攝取葡萄糖達到 正常細胞二至十倍，造成掃描圖像上出現明顯的 「光點」  能於癌細胞的早期 ( 約 0.5 公分 ) 準確地判定癌細 胞，提供醫師作為診斷及治療的依據，診斷率高 達 87-91 ％， 30 歲以上的成年人及有癌症家族史 的民眾，建議每隔 1 ～ 2 年做一次 PET 檢查。

20 PET (Positron emission tomography)

21

22 Cell identification via microscope  Tools Traditional optical microscope  Stained specimen Fluorescent microscope  Identification for nuclear and gene expression Laser confocal microscope  Identification from 2-D to 3-D Multi-photon microscope  Identification from 2-D to 3-D

23 Applications in a hospital  Assist surgeon plan surgical operation or diagnose  Picture archiving system (PACS) 將醫療系統中所有的影像，以數位化的方式儲存，並經 由網路傳遞至同系統中，供使用者於遠側電腦螢幕閱讀 影像並判讀。  Telemedicine  Surgical simulation: Medical Visualization ， Surgical augmented Reality, Medical- purpose robot, Surgery Simulation ， Image Guided Surgery ， Computer Aided Surgery  Estimate the location, size and shape of tumor

24 PACS System

25 Virtual Surgery

26

27 Related techniques  Classification of normal or abnormal tissues such as carcinoma Pre-processing: Contrast enhancement, noise removal, and edge detection Lesion segmentation: extract contours of interest  thresholding  2-D segmentation  3-D segmentation based on voxel data  Color image processing

28 Our study  Virtual colonoscopy  Bone tumor segmentation with MRI and virtual display  Breast carcinoma based on histology and cytology  Visualization of cell activities using confocal laser scanning microscope

29 Virtual colonscopy-Browsing or navigation within a colon  Helical CT – patients injected contrast medium  Re-sampling — Voxel-based  Interpolation  Automatic segmentation (seed) threshloding  Determination of the skeleton of the colon  Connected-Component Labeling  Surface rendering and volume rendering  Extraction of suspicious sub-volumes for diagnosis

30

31 Automatic segmentation

32 Determination of the skeleton of the colon

33 Display and measurement

34 Bone tumor segmentation with MRI and virtual display—Contrast medium  Otsu thresholding Region growing  Tri-linear interpolation  Morphological post-processing Morphological post-processing  Surface rendering  Measurement

35 Histogram of T1 weighted and T2 weighted

36

37

38 (a) 0 度 (b) 45 度

39 Classification of Breast Carcinoma

40

41

42 正常異常 系統判斷為正常 126 系統判斷為異常 111 準確性 敏感度有效性 76.67%64.71%92.31%

43

44 應用雷射共軛焦顯微鏡影像三維重構研究螢 光細胞之活動 - 以子宮頸癌細胞為例

45 細胞結構簡介

46 雷射共軛焦顯微鏡之成像原理 46 雷射共軛焦顯微鏡解析度 :

47 雷射共軛焦顯微鏡雜訊生成之原因  大部分的生物樣本，潛在一些特性會降低 CLSM 影像解析度： 第一個是具有散射的特性 第二個特性是折射率的不匹配 (refractive- index mismatch) 所產生的。  由於折射率的不匹配會引入球面像差，而使得縱 向與橫向解析度變差。 47

48 散射的特性 48 散射光強度 :

49 散射的特性 49

50 研究影像 50

51 實驗方法與架構 51 選取重要影像 重要影像初始輪 廓偵測 Snake 自動偵測 重要影像輪廓 利用適應性 Snake 自動偵測 整體輪廓 排除對比較差的 細胞核輪廓 重建三維細胞 分割蛋白質劇烈 活動區域 二質化細胞質 區域

52 選取重要影像  由於雷射共軛焦取像環境的限制，在細胞邊界處通常訊號 較弱且較為模糊，使得初始輪廓分割相當困難，因此本步 驟選取出對比最好的影像作為重要影像 ( 分割時的初始切 片 ) ，偵測出其初始輪廓。 52

53 重要影像初始輪廓偵測 53

54 Snake 自動偵測重要影像細胞輪廓 - 去除人工雜物  由於雷射共軛焦取像的特性，其邊緣部分通常模糊不清， 因此初始輪廓的結果偶爾會產生過小得輪廓，本研究排除 輪廓長度小於 100pixel 得輪廓 ( 可能是雜訊的輪廓 ) ，僅以較 大的輪廓作為初始輪廓。 54

55 利用適應性 Snake 自動偵測整體影像輪廓 55

56 利用適應性 Snake 自動偵測整體影像輪廓 56

57 輪廓重疊偵測 - 濾除對比較差細胞核輪廓 偵測  細胞核訊號分佈模式 : 重要影像前半部 重要影像後半部 57 = 細胞質區域 Mean = 細胞核區域 Mean

58 輪廓重疊偵測 - 濾除對比較差細胞核 輪廓偵測 58

59 59

60 60

61 細胞內蛋白質反應劇烈區域的分割 61

62 細胞內蛋白質反應劇烈區域的分割  初始點決定與 K-Means 群法偵測最亮區域 62

63 不同角度顯示 63

64 其他範例 (Case3) 64

65 不同角度顯示 65

66 三維重建資料比較 66

67 體積比例量測 Case 1Case 2Case3 細胞質區域 (Voxels) 224396521563629562 蛋白質活動劇烈區域 (Voxels) 381933877785 比例 1.7019%0.649%1.236% 67

68 效能評估 ProcessCase1:Time(Sec)Case2:Time(Sec)Case3:Time(Sec) 細胞質區域分割 687483 蛋白質活動劇烈區 域分割 767 整體三維重建 9109 整體時間 849099 68

69 拉普拉斯三維平滑 69

70 拉普拉斯三維平滑 70

71 Requirements for medical image processing system in clinical diagnosis  Automatic and less human interaction  Qualitative and quantitative measurements  Stable and reliable (experiments with much more cases)  Performance evaluation True positive, true negative, false positive, false negative Accuracy, sensitivity, and specificity Receiving operating characteristic curve (An index for evaluating the effectiveness of classification  Optimal classification threshold  Area under ROC approach 1 – better classification

72 ROC curve

73 Analyses of prognosis on breast cancer for a stained tissue  Microscopy with different resolution (400 or 100) for a stained tissue  Fluorescent microscopy in detecting the number of chromosome  Immunohistochemistry(IHC)

74 Her-2 IHC image

75 Fish image(normal)

76 Fish image (abnormal)

77 Preliminaries or problems ?  Blurring often caused by patient motion or respiration  Clinical opinion or idea obtained from an experienced surgeon  Non-absolute answers at some specific conditions  Trade-off between complexity and performance  Large variations for different image modality

78 Preliminaries or problems ?  Automation is necessary so as to help physicians  Prove identification accuracy — comparison between manual and image processing approaches  Classification based on neural network, pattern recognition, or fuzzy,.. etc is crucial in practical applications

79  Thanks for your attention!