Introduction Segmentation Implementations and Results Summary Automatic segmentation of liver tumors from MRI images. Rune Petter Sørlie Department of Physics University of Oslo thesis presentations, November 2005 Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Outline 1 Introduction Treating liver tumors Guidance of probes Monitoring the process 2 Segmentation Preprocessing Thresholding and edge detection Methods Evaluation of segmentation 3 Implementations and Results Segmentation Results Discussion Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Introduction Want an automatic segmentation of liver tumors. Support Laparascopic procedures. Utilize the Open MR-scanner at IVS. Improve patient treatment. The thesis present three segmentation methods Snakes Level set methods Watershed segmentation Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Magnetic Resonance imaging (MRI) A historical perspective 1946 - Nuclear magnetic resonance discovered. 1972 - First image found place. 1986 - Norway bought their first MR-scanners. MRI advantages Excellent contrast detail of soft tissue. Image caption in any plane. Easy image interpretation. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Methods to destroy a tumor. We will show two probe based methods which allow minimum invasive treatment. Cryoablation RF ablation Pros and cons Less risk of infections. Patients will restore to health sooner. Not suitable for tumors >3cm. High rate of metastases reoccurances (Cryo). Fails to damage cells near major vessels. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Cryoablation Cryoablation uses liquid nitrogen to freeze the tisse surounding the probe. At about -40◦ C the cell membrane cracks and the cells are damaged. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Cryoablation Typical tumor is 2-3cm and a rim of 1cm around the tumor is frozen. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process RF ablation Makes use of Radio Frequency alternating current to heat up tissue. Cures tumors <3cm. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Guidance and probes Both methods make use of probes. The surgeon want guidance when inserting the probe into the tumor. Different modalities are used A combination of CT and ultrasound Fluoroscopy (X-ray) IVS use the Open MR scanner. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Monitoring ablation We will search to optimize the ratio of damaged cells and tumor cells. Must eliminate all tumor cells Minimize ablation of healthy cells Possible by thermal monitoring Makes use of a thermal model Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Principal of RF ablation Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Diagnostic image/pre surgery Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process Guidance of the RF-probe Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process 24 hours after treatment Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Treating liver tumors Guidance of probes Monitoring the process 9 months later, complete treatment Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation The need for preprocessing Adjust for variable background light Estimate background by smoothing Subtract estimated background from input image Dealing with Artifacts Echoes from vessels Patient movements –> blurred image Best handled by a new image Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation cont. The need for preprocessing Tissue variations looks like noise. The smoothing dilemma Remove speckle Conserve edges Smoothing approaches Degree of smoothing Size and orientation of filter kernel Median/Mean/Gauss filtering Iterated smoothing Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Thresholding Gray level thresholding Basic approach Objects presented as regions Neighborhood information ignored Local threshold, Canny filter Adaptive to local gray level Finds to many edges or... Does not make a closed curve Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Connected edge Finding a connected edge by using contours Snakes Level set methods Region based segmentation Watershed segmentation Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Snakes, an active contour Described as a parametric curvature v (s) = [x(s), y (s)]. The snake will fit into the image with use of energy minimization. ∗ Esnake Z 1 {Eint (v (s)) + Eimage (v (s)) + Econ (v (s))}ds = (1) 0 Z Eint = 0 1 1 (α|v 0 (s)|2 + β|v 00 (s)|2 ) 2 Eimage = −|∇I(x, y )|2 Rune Petter Sørlie Segmenting liver tumors (2) (3) Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation GVF snakes Snake steered with Gradient Vector Flow. To minimize energy, must satisfy the Euler equation αx 00 (s) − βx 0000 (s) − ∇Eext = 0 (4) written as a force balance equation (p) Fint + Fext = 0 (p) where Fint = αx 00 (s) − βx 0000 (s) and Fext = −∇Eext Rune Petter Sørlie Segmenting liver tumors (5) Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Let snake function x be a function of s and time, x(s, t) Sets the partial derivative of x, xt (s, t) = ∂x/∂t to zero. Then equation 6 gives the snake position for equilibrium. xt (s, t) = αx 00 (s) − βx 0000 (s) − ∇Eext (6) Introduces a computational diffusion on homogeneous areas. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation GVF snake Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Principle of the Level set method Based on moving curvature. Adds one dimention to the problem domain Parameterizing har limitations Corners Change in topology Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Level set The propagation is described by a speed function F which can be a function of curvature κ. κ= gxx gy2 − 2gxy gx gy + gyy gx2 (gx2 + gy2 )3/2 Rune Petter Sørlie Segmenting liver tumors (7) Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Level set The propagation is described by a speed function F which can be a function of curvature κ. κ= gxx gy2 − 2gxy gx gy + gyy gx2 (gx2 + gy2 )3/2 Rune Petter Sørlie Segmenting liver tumors (8) Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation The Threshold Level Set Method P(x) = g(x) − L , g(x) ≤ C U − g(x) , g(x) > C where C = Rune Petter Sørlie U −L +L 2 Segmenting liver tumors (9) Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Threshold Level set segmentation Propagation term P Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Watershed segmentation Watershed segmentation simplified to one dimension Terrain model, slow and inaccurate Bottom-up approach Oversegmentation solved with postprocessing (hierarchy, min. depth) Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation The need for a referance Evaluate the segmentation against the truth? A gold standard Manual segmentation Synthetic images Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Preprocessing Thresholding and edge detection Methods Evaluation of segmentation Dice Similarity Coefficient (DSC) DSC usefull for messure of spatial overlap. For two regions A and B, DSC is given by DSC(A, B) = 2(A ∩ B) A+B no overlap, DSC=0 full overlap, DSC=1 Rune Petter Sørlie Segmenting liver tumors (10) Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Snake Segmentation with GVF snake Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Snake same image, other initialization Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion frametitleSnake Detail of GVF field Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Level set segmentation Fastmarching level set segmentation Initialized to (135, 135), σ = 1.0, α = −0.20, β = 4.0, threshold = 100, stopvalue = 100. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Watershed segmentation Iterations=5, conductance=3.0, lower threshold=0.01, max depth=0.20 Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Level set segmentation in 3D, Exam 2212 Threshold level set: seedpoint = (98, 137, 14), radius = 12, thresholds: L = 50, U = 70. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Level set segmentation in 3D seedpoint = (140, 167, 9), radius = 5, thresholds: L = 0, respectively U = 58 and U = 55 Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Segmentation Results Discussion Interpreting the results method snake threshold level set watershed parameters r= 4.0 r= 6.0 r=10.0 r=14.5 u=135 u=140 u=145 t=0.01, d=0.20 t=0.001, d=0.20 t=0.02, d=0.18 DSC 0.756 0.916 0.944 0.903 0.902 0.916 0.899 0.896 0.896 0.896 time (s) 1.24 1.32 1.17 1.18 1.36 2.23 1.31 10.20 11.20 9.80 r=radius, u=upper threshold, t=lower threshold and d=max. depth. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Summary Most liver tumors possible to segment automatic. Quality of segmentation close to manual segmentation. Until recently ITK is used to segment organ and their subparts. Time has come to include tumor segmentation in ITK based applications. Rune Petter Sørlie Segmenting liver tumors Introduction Segmentation Implementations and Results Summary Questions? Rune Petter Sørlie Segmenting liver tumors