1
Computer Aided Diagnosis in Breast Ultrasound Imaging
Ruey-Feng Chang (張瑞峰)
Department of Computer Science and Information Engineering (資訊系)
Graduate Institute of Biomedical Electronics and Bioinformatics (生醫電資所) National Taiwan University(國立台灣大學), Taiwan
Yi-Hong Chou
(周宜宏)
Veterans General Hospital, Taiwan
Chiun-Sheng Huang (黃俊升)
National Taiwan University Hospital, Taiwan
Yeun-Chung Chang (張允中)
National Taiwan University Hospital, Taiwan
Jeon-Hor Chen (陳中和)
University of California, Irvine, USA
Woo Kyung Moon
Seoul National University Hospital, Korea
Etsuo Takada
Dokkyo University, Japan
2
Contents
2-D Breast US CADx
Elastography CADx
3-D Breast US/ABUS CADx
ABUS CADe
ABUS/MRI Density Analysis
MRI CADe/CADx
3
NTU CAD Lab
1997
Aloka SSD 1200
2003
Acuson, GE, Medison, ATL, Philips
1999
Free-hand 3D US
2000 Voluson 530
2002 GE 730
2010 GE E8 and E6
2-D US 1997
3-D US
2000 2002
Whole Breast
2002 Free-hand
2003 Aloka ABUS
2005 U-systems
2004
2004
Telemed, Lithuania Echo Blaster 128
2005
Terason t3000
PC-based US
Elastography
2006
MRI
2008 2011
Automated US
2006
Hitachi Elastography
2010
Siemens Elastography
2011
Siemens ABVS
2011, 2013
2D/3D Shearwave
2008
Morphology, Kinetic
2009 Tofts Model
2010 DWI
2010 Siemens ARFI
2011 Supersonic Shearwave
20133D Elastography
Introduction
For 2D breast US, the physician has detected the tumor and only the
computer-aided diagnosis (CADx) is needed for the tumor.
For the new automated whole breast US (ABUS), the computer-aided
detection (CADe) is needed for detecting the tumors, just like mammography CAD.
4
Technology Development Program for Academic (TDPA, 學界科專)
This TDPA project was supported by the Ministry of Economic Affairs
(MOEA) to develop
– a CADe system for ABUS – a CADx system for B-mode
US/elastography
– breast US GPS/recoding System
The Co-PIs are Dr. Chou from VGH, Dr. Huang, and Dr. Chang from NTUH.
8 PhD students worked on this
project.
5Technology Development Program for Academic (TDPA, 學界科專)
Two patents have been applied.
– Breast Ultrasound Scanning and Diagnosis Aid System
– Ultrasound Imaging Breast Tumor
Detection and Diagnostic System and Method
25 international journals and 12 international conference papers
– Three IEEE Trans. MI papers have been published.
The CADe and CADx systems have
been transferred to TaiHao Medical Inc.
(http://taihaomed.com/)
6
CADx for TDPA
8
CADx for TDPA
ABUS CADe
9
DEMO
10
DEMO
11
Breast US GPS/Recoding System
12 Tracking map Captured US video
Tumor location Probe direction
2-D BREAST US CAD
- ROBUST TEXTURE ANALYSIS
“Computer-aided Diagnosis Applied to US of Solid Breast Nodules by Using Neural Networks”, Radiology, vol. 213, no. 2, pp.407-412, 1999.
“Robust texture analysis using multi-resolution gray-scale invariant
features for breast sonographic tumor diagnosis,” IEEE Transactions on Medical Imaging, vol. 32, no. 12, pp. 2262-2273, 2013.
“Computer-aided diagnosis for distinguishing between triple-negative breast cancer and fibroadenomas based on ultrasound texture features,”
Medical Physics, vol. 42, no. 6, pp. 3024-3035, 2015.
13
14
2D US CAD
“Computer-aided Diagnosis Applied to US of Solid Breast Nodules by Using Neural Networks”, Radiology, vol. 213, no. 2, pp.407-412, 1999.
Editorial: Georgia D. Tourassi, "Journey toward Computer-aided Diagnosis: Role of Image Texture Analysis,” Radiology, vol. 213, no.2, pp.317-320, Nov. 1999.
Robust Texture Analysis
Because the parameters of ultrasonic machine is adjustable, the images from the same machines may have different texture information.
Moreover, the images from different ultrasonic machines have different texture information.
Hence, the ranklet transform is
proposed in this study to extract the
gray-scale invariant texture features for tumor diagnosis.
“Robust texture analysis using multi-resolution gray-scale invariant features for breast 15
sonographic tumor diagnosis,” IEEE Transactions on Medical Imaging, vol. 32, no. 12, pp.
2262-2273, 2013.
Ranklet Transform
Similar to wavelet transform, the ranklet images can be derived from a family of multi-resolution, orientation-selective features
Differently from wavelet transform, it deals with ranks of pixels rather than with their gray-scale intensity values
16 HH (Diagonal subband, D)
HL (Horizontal subband, H) LH (Vertical subband, V)
Ranklet Example
The ranklet images from images with non-linear monotonic change filters (i.e., gamma correlation, histogram equalization) are nearly the same.
17
Origin Wavelets Ranklets
W-1V W-1H W-1D R8V R8H R8D
W-1V W-1H W-1D R8V R8H R8D
Proposed Robust Texture Analysis
First, we decompose each BUS image from the test database into ranklets.
Afterwards, the gray-scale invariant texture features based on GLCM can be calculated from each
transformed image with regions of interest (ROIs).
The SVM classifier is adopted to distinguish a benign tumor from a malignant one.
18
Input BUS Data GLCM Texture
Feature Extraction
Training Phase
Testing Phase
Ranklet Transform Ranklet Decomposition
I0
H 1
R- R-V1
D 1
R-
1
R-
Texture Feature Representation
SVM Training/Testing
Input Database SVM
Training/Testing
Three US Machines
Database A includes 116 subjects (78 benign and 38 malignant cases) obtained with Acuson Sequoia
machine.
Database B includes 193 subjects (133 benign and 60 malignant
cases) obtained with GE LOGIQ 7 machine.
Database C includes 161 subjects (104 benign and 57 malignant
cases) obtained with GE Voluson 730 expert machine.
19
Experiments
The GLCM-based textural features are applied for original image (origin), multi- resolution wavelet images (wavelets)
and multi-resolution ranklet images (ranklets).
20
D.B. Method AUC ACC (%) SENS (%) SPEC (%)
A
Origin 0.81±0.03 74.28±2.27 63.93±5.78 79.39±3.14 Wavelets 0.84±0.03 79.45±1.73 70.54±4.23 83.76±2.39 Ranklets 0.90±0.02 81.68±1.69 69.66±4.63 87.55±2.15
B
Origin 0.86±0.03 78.75±2.21 66.53±5.99 84.33±2.89 Wavelets 0.92±0.02 84.23±1.71 74.99±4.74 88.38±2.12 Ranklets 0.94±0.02 86.35±1.64 79.56±4.44 89.35±2.11
C
Origin 0.84±0.03 76.49±2.28 67.74±5.30 81.31±3.32 Wavelets 0.85±0.02 77.14±1.74 69.27±3.87 81.48±2.51 Ranklets 0.92±0.02 84.58±1.70 81.50±3.66 86.19±2.44
Cross-machine Experiments
The AUC of cross-platform training and testing
21
Train
Database Method Test Database
A
B C B+C
Origin (95% CI)
0.783 (0.717-0.846)
0.823*
(0.769-0.892) 0.786 (0.739-0.835) Wavelets
(95% CI)
0.722†
(0.6283-0.788) 0.752†
(0.663-0.816) 0.729†
(0.679-0.782) Ranklets
(95% CI)
0.934*
(0.896-0.965) 0.877
(0.825-0.929)
0.876*
(0.837-0.909)
B
A C A+C
Origin (95% CI)
0.724 (0.633-0.816)
0.807*
(0.736-0.869) 0.764†
(0.707-0.817) Wavelets
(95% CI)
0.757†
(0.643-0.828) 0.842*
(0.784-0.906) 0.832*
(0.779-0.879) Ranklets
(95% CI)
0.867 (0.792-0.929)
0.873 (0.817-0.922)
0.875*
(0.825-0.909)
C
A B A+B
Origin (95% CI)
0.709†
(0.614-0.806) 0.789
(0.720-0.860)
0.765†
(0.708-0.821) Wavelets
(95% CI)
0.795 (0.677-0.864)
0.785†
(0.712-0.854) 0.808†
(0.731-0.843) Ranklets
(95% CI)
0.859 (0.780-0.913)
0.913*
(0.871-0.949) 0.891*
(0.855-0.925)
TNBC vs. Fibroadenoma
Triple-negative breast cancer (TNBC) is frequently misclassified as fibroadenoma due to benign
morphologic features on breast ultrasound.
The multi-resolution ranklet features are proposed to to discriminate between TNBC and benign
fibroadenomas.
22
1cm benign fibroadenoma 1cm triple-negative breast cancer
“Computer-aided diagnosis for distinguishing between triple-negative breast cancer and fibroadenomas based on ultrasound texture features,” Medical Physics, vol. 42, no. 6, pp.
3024-3035, 2015.
Experiments
169 tumors, including 84 benign fibroadenomas and 85 TNBCs, are used in this study
23
Accuracy (%) Sensitivity (%) Specificity (%) PPV (%) NPV (%) Az
Morphology 76.92*
(130/169)
78.82* (67/85)
75.00* (63/84)
76.13* (67/88)
77.78*
(63/81) 0.8470*
Conventional Texture (GLCM)
79.29* (134/169)
81.18* (69/85)
77.38* (65/84)
78.41* (69/88)
80.25*
(65/81) 0.8542*
Invariant Ranklet texture
87.57 (148/169)
89.41 (76/85)
85.71 (72/84)
86.36 (76/88)
88.89
(72/81) 0.9695
Combined 93.49
(158/169)
94.12 (80/85)
92.86 (78/84)
93.02 (80/86)
93.98
(78/83) 0.9702
ROC Curves
24
Elastography CADx
- SHEAR-WAVE ELASTOGRAPHY
“Analysis of elastographic and B-mode features at sonoelastography for breast tumor classification”, Ultrasound in Medicine and Biology, vol. 35, no. 11, pp. 1794–1802, 2009.
“Automatic selection of representative slice from cine-loops of real-time sonoelastography for classifying solid breast masses”, Ultrasound in Medicine and Biology, vol. 37, no. 5, pp. 709-718, 2011.
“Breast Tumor Classification Using Fuzzy Clustering for Breast
Elastography”, Ultrasound in Medicine and Biology, vol. 37, no. 5, pp.
700-708, 2011.
“Classification of breast tumors using elastographic and B-mode
features: Comparison of automatic selection of representative slice and physician-selected slice of images”, Ultrasound in Medicine and Biology, vol. 39, no. 7, pp. 1147-1157, 2013.
25
Dynamic Stain Elastography
A lot of elastography slices are obtained during the operator to compress a tumor.
– That is, the stain elastography image is dynamic.
26
Benign Malignant
Elastography CAD
Benign or Malignant
Elastographyfeatures B-modefeatures
Selecting representative slice
Segmentation
A representative slice of the dynamic elastography image needs to be
selected.
The B-mode image could be used to segment the tumor and the tumor
contour could be applied to the corresponding
elastography image.
Both the elastography features and B-mode
features could be used to diagnose the tumor.
“Analysis of elastographic and B-mode features at sonoelastography for breast tumor classification”, Ultrasound in Med. & Biol., Vol. 35, No. 11, pp. 1794–1802, 2009.
27
Image Quality Quantification
Signal to Noise Ratio (SNR) and Contrast to Noise Ratio (CNR) are used to quantify the elastography image quality.
– The middle block in the tumor is used to calculate the SNR.
28
Low SNR High SNR
SNR mean
standard_deviation
“Automatic selection of representative slice from cine-loops of real-time sonoelastography for classifying solid breast masses”, Ultrasound in Med. & Biol., Vol. 37, No. 5, pp. 709-718, 2011.
Image Quality Quantification
In the CNR method, not only the middle
block but also two outside blocks are used.
Low CNR High CNR
29Over-compressed
2
2 2
( )
_
middle_mean outside_mean CNR middle_SD outside SD
The Strain Curve
30
SNR CNR
Tumor Segmentation
The contrast-enhanced gradient
image is used to segment the tumor using the level set method.
Gradient Magnitude Sigmoid
Sigmoid Level Set
Result
31
BI-RADS Features
Shape
– Oval, Round, Irregular
Orientation
– Parallel, Not parallel
Margin
– Circumscribed
– Not circumscribed
• Indistinct, Angular, Microlobulated, Spiculated
Lesion boundary
– Abrupt interface, Echogenic halo
Echo pattern
– Anechoic, Hyperechoic,
Complex, Hypoechoic, Isoechoic
Posterior shadowing
– No posterior acoustic features, Enhancement, Shadowing,
Combined patter
32BI-RADS Atlas
Elastography Features
Stiffness Ratio
Elasticity mean
Lesion boundary elasticity
33
Experimental Results
The data consisted of 151 biopsy-proved lesions (89 benign and 62 malignant lesions) from Dr. Moon.
For CNR, elastography (82.12%) is better than B-mode (80.79%).
For SNR and Physician-selected, B-mode (87.42%, 84.11%) is better than elastography (82.12%, 82.78%).
SNR (90.07%) is better than CNR (86.09%) .
SNR (90.07%) is similar to Physician-selected (89.40%) .
34
Features Accuracy Sensitivity Specificity PPV NPV
CNR
B-mode 80.79% 70.97% 87.64% 80.00% 81.25%
Elastography 82.12% 74.19% 87.64% 80.70% 82.98%
All Features 86.09% 82.26% 88.76% 83.61% 87.78%
SNR
B-mode 87.42% 83.87% 89.89% 85.25% 88.89%
Elastography 82.12% 79.03% 84.27% 77.78% 85.23%
All Features 90.07% 90.32% 89.89% 86.15% 93.02%
Physician selected
B-mode 84.11% 77.42% 88.76% 82.76% 84.95%
Elastography 82.78% 74.19% 88.76% 82.14% 83.16%
All Features 89.40% 85.48% 92.13% 88.33% 90.11%
“Classification of breast tumors using elastographic and B-mode features: Comparison of automatic selection of representative slice and physician-selected slice of images”, Ultrasound in Medicine and Biology, vol. 39, no. 7, pp. 1147-1157,2013
ROC curves for SNR Representative Slice
Elastography, Az=0.8461 B-mode, Az=0.9223
B-mode + Elastography Az=0.9407
35
True Negative Case
CNR SNR
Physician-selected
Fibroadenomas
36True Positive Case
CNR SNR
Physician-selected
Infiltrating carcinomas
37Summary
The performance of B-mode features is better than that of elastography features in this study.
Combining the B-mode and
elastography features could improve the diagnosis performance.
The proposed image quantification methods could have similar
performances with the physician.
– The proposed selection of representative slice is robust and reliable.
38
Shear-wave Elastography
Stain elastography requires manual tissue compression which is operator dependent.
Shear-wave elastography (SWE) is a new method to permit absolute quantification of tissue stiffness.
Using acoustic radiation to stimulate tissues
39
Benign Softer (low kPa)
Malignant Harder (high kPa)
Shear-wave CAD
40
Shear wave
image Segmentation
B-mode features
Elastographic features Seed
Diagnosis
Result
Experimental Results
SuperSonic Imagine Aixplorer US
109 breast tumors
– 57 benign and 52 malignant
41
109 cases: 57 benign and 52 malignant
Accuracy (%) Sensitivity (%) Specificity (%) PPV (%) NPV (%) Az
B-mode 87.16
(95)
86.54 (45)
87.72 (50)
86.54 (45/52)
87.72
(50/57) 0.8312
Elastography 89.91 (98)
86.54 (45)
92.98 (53)
91.84 (45/49)
88.33
(53/60) 0.9511
Combined 94.50
(103)
92.31 (48)
96.49 (55)
96.00 (48/50)
93.22
(55/59) 0.9705
ROC Curves
42
3D/4D Elastography
The new CAD systems could be further developed for the 3D/4D elastography to provide more robust diagnosis.
43
From Dr. Takada From Dr. Moon and Dr. Chou
3-D BREAST US/ABUS CADX
“Characterization of Spiculation on Ultrasound Lesions”, IEEE Transactions on Medical Imaging, vol. 23, no. 1, pp. 111-121, 2004.
“Solid Breast Masses: Neural Network Analysis of 3-D Power Doppler Ultrasound Image Features for Classification as Benign or Malignant”, Radiology, vol. 243, no. 1, pp. 56-62, 2007.
“Analysis of Tumor Vascularity Using Three-Dimensional Power Doppler
Ultrasound Images,” IEEE Transactions on Medical Imaging, vol. 27, no. 3, pp.
320-330, 2008.
“Vascular Morphology and Tortuosity Analysis of Breast Tumor Inside and Outside Contour by 3-D Power Doppler Ultrasound”, Ultrasound in Med. &
Biol., vol. 38, no. 11, pp. 1859-1869, 2012.
“Computer-aided diagnosis for the classification of breast masses in automated whole breast ultrasound images,” Ultrasound in Medicine and Biology, vol. 37, no. 4, pp. 539-548, 2011.
44
45
3D US Spiculation
A kind of stellate distortion caused by the intrusion of the breast cancer.
The spiculation displayed only on the C- view of 3D US.
– The conventional 2D US cannot see the spiculation.
*IEEE Transactions on Medical Imaging, vol. 23, no. 1, pp. 111-121, Jan. 2004. 45
Tumor
Spiculation
46
Experimental Result
Spiculations are more easily seen in malignant tumors.
Spiculations are more likely to appear in the coronal view.
0 40 80
1 120
(a) Detection slices
Dataset no.
0 40 80
1 120
(b) Detection slices
Dataset no.
0 40 80
1 120
(c) Detection slices
Dataset no.
Longitudinal view
0 40 80
1 105
(b) Detected slices
Dataset no.
0 40 80
1 105
(c) Detected slices
Dataset no.
0 40 80
1 105
(a) Detected slices
Dataset no.
Coronal view Transverse view
120 benign 3D datasets
105 malignant 3D datasets
3D Power Doppler US
A useful tool to detect the vessels.
– 2D Doppler US can not obtain the entire 3-D vessels.
The blood supply and vessel distribution are the important features to diagnose an tumor.
“Solid Breast Masses: Neural Network Analysis of 3-D Power Doppler Ultrasound Image 47
Features for Classification as Benign or Malignant”, Radiology, vol. 243, no. 1, pp. 56-62, 2007.
48
System Overview
Preprocessing
Feature Extraction
Neural Network 3-D
Thinning Algorithm Building
Vascular Tree Original Dataset
Malignancy Prediction vascular points
one-pixel wide skeleton
Thinning Result
Vascular Tree Construction
To generate a skeleton representation, the thinning result was converted into tree structure using breadth first search (BFS) algorithm.
Bifurcate node root
x
y
z
●
●
●
●
○
○ ○ ○
○
○
○ ○ ○
root
x
y
z
●
●
●
●
○
○ ○ ○
○
○
○ ○ ○
51
Feature Extraction
Ten features are used for the vessel tree.
– Vessel-to-Volume Ratio (R
v)
– Number of Vascular Trees (N
v) – Measurement of Length
• Total length (L
1)
• Length of the longest path (L
2)
– Bifurcation (Bn) – Diameter (D
v)
– Number of cycles (NC)
– Tortuosity measures (DM, ICM, SOAM)
52
Benign Case
Original Thinning
All Vessels
Longest
Vessel
53
Malignant Case
Original Thinning
All Vessels
Longest
Vessel
54
Benign vs. Malignant
The bottom images are the longest vessel.
The longest vessel of malignance has more curvature.
Benign Malignant
55
Vascular Analysis Inside/outside Tumor
To evaluate morphologic and tortuous
features of vessels inside and outside the tumor region on 3-D power Doppler US
“Vascular Morphology and Tortuosity Analysis of Breast Tumor Inside and Outside Contour 55
by 3-D Power Doppler Ultrasound”, Ultrasound in Med. & Biol., vol. 38, no. 11, pp. 1859-1869, 2012.
Benign
56
Malignant
The significant vessel trees outside (red) and inside (white) tumor.
The tumor contour (blue) and primary path outside tumor (red) and inside tumor (white)
2-D PDUS Vessels (orange) and skeletons
(red and white)
ROC Curves
There were 113 solid breast masses (60 benign, 53 malignant).
57
ABUS CADx
2D/3D CADx could be applied for ABUS.
58 3-D breast tumor VOI
Level set segmentation
B-mode image Tumor mask
Texture feature Shape feature Ellipsoid fitting feature
Classification using binary logistic regression
“Computer-aided diagnosis for the classification of breast masses in automated whole breast ultrasound images,” Ultrasound in Medicine and Biology, vol. 37, no. 4, pp. 539-548, 2011.
3D Tumor Segmentation
Benign
Malignant Example
60
147 cases (76 benign and 71
malignant breast masses) were obtained by U-systems ABUS.
61 GLCM Shape Ellipsoid
fitting
GLCM and shape
GLCM and ellipsoid
fitting
Shape and ellipsoid
fitting
ALL
Accuracy (%) (111/147)75.51 (121/147)82.31 (117/147)79.59 (118/147)80.27 (120/147)81.63 (125/147)85.03 (121/147)82.31 Sensitivity (%) (58/71)81.69 (60/71)84.51 (54/71)76.06 (59/71)83.10 (60/71)84.51 (60/71)84.51 (59/71)83.10 Specificity (%) (53/76)69.74 (61/76)80.26 (63/76)82.89 (59/76)77.63 (60/76)78.95 (65/76)85.53 (62/76)81.58 PPV (%) (58/81)71.60 (60/75)80.00 (54/67)80.60 (59/76)77.63 (60/76)78.95 (60/71)84.51 (59/73)80.82 NPV (%) (53/66)80.30 (61/72)84.72 (63/80)78.75 (59/71)83.10 (60/71)84.51 (65/76)85.53 (62/74)83.78
Az 0.8603 0.9138 0.8496 0.9153 0.8195 0.9466 0.9388
ABVS CADe
- TUMOR DETECTION - TUMOR MAPPING
“Computer-Aided Tumor Detection Based on Multi-Scale Blob
Detection Algorithm in Automated Breast Ultrasound Images”, IEEE Transactions on Medical Imaging, vol. 32, no. 7, pp. 1191-1200, July 2013.
“Tumor detection in automated breast ultrasound images using quantitative tissue clustering”, Medical Physics, vol. 41, no. 4, pp.
042901-1-8, April 2014.
“Multi-dimensional tumor detection in automated whole breast ultrasound using topographic watershed,” IEEE Transactions on Medical Imaging, vol. 33, no. 7, pp. 1503-1511, July 2014.
“Feasibility Testing: Three-dimensional Tumor Mapping in Different Orientations of Automated Breast Ultrasound,” Ultrasound in
Medicine and Biology, vol. 42, no.5, pp. 1201-1210, 2016. 62
Automated Whole Breast US
63
Aloka, U-systems (acquired by GE), and Siemens, SonoCiné (distribution agreement with Philips), and iVu have developed the automated whole breast ultrasound (ABUS) systems.
Aloka, 2006.07
Aloka
U-systems
U-systems, 2006.09
Siemens
SonoCiné
iVu
Siemens ABVS
ACUSON S2000 Automated Breast Volume Scanner (ABVS)
Siemens: 736×481×318 = 0.208mm×0.052mm×0.526mm
ABVS Viewer
Our ABVS view system has been
transferred to TaiHao Medical Inc, Taiwan.
– BR-ABVS Viewer 1.0
– FDA cleared, 2016.01
– http://taihaomed.com/
ABVS Viewer (2)
Our ABVS view system is applied for a ABUS system from a Taiwan company.
– The probe is rotated to obtain a whole breast image in only one scanning.
– Begin clinical trials in NTUH and other hospitals
Cooperation with China company for CFDA
66
Supine/Prone ABUS
Supine ABUS
– Same position with surgery
– At least three passes for a breast
Prone ABUS
– Same position with MRI – One pass for a breast
Reconstruction of rotated prone ABUS (iABUS/iVu)
67 Inferior
Lateral
AP Medial
Superior
iABUS Viewer
68
2.9 clock / 34 mm (123.9.91) Transverse view
87 degree Radial view Sagittal view
Coronal view
Dual-modality ABUS System
Combines an FFDM X-ray machine with automated breast ultrasound (ABUS) technology
– CapeRay Medical Ltd.
– A long probe is scanned under the plate.
69
ABUS/MAMMO Viewer
70
Mammo
Watershed Tumor Detection
Our ABUS CADe is based on a watershed transform method.
– The watershed transform was applied to
gather similar tissues around local minima to be homogeneous regions.
This method detects every tumor, but
some non-tumors (false positive, FP) are also detected.
– Hence, the likelihoods of being tumors of the regions were estimated using the quantitative morphology, intensity, and texture features in the 2-D/3-D false positive reduction (FPR).
“Multi-dimensional tumor detection in automated whole breast ultrasound using topographic 71
watershed,” IEEE Transactions on Medical Imaging, vol. 33, no. 7, pp. 1503-1511, July 2014.
Proposed System
72
ABUS image
Resolution reduction &
image enhancement Watershed transform for tissue segmentation
Suspicious abnormality extraction
2-D/3-D false positive reduction
Slice by slice processing
False Positive Reduction (FPR)
2013/7/27
73 Original
image
Watershed transform
Abnormality extraction
2-D FPR
3-D FPR
Experiments
ABUS SomoVu ScanStation
– 138 cases (104 abnormal and 34 normal) – 104 breast lesions of 104 patients
• 68 benign lesions
• 65 malignant lesions
• Lesion size : 1.75±1.13 cm
10-fold cross validation (10F-CV) is used.
FROC curves and the jackknife alternative of FROC-1 (JAFROC) figure of merit (FOM)
were performed to evaluate the performance of the CADe system
74
FPs/pass after applying 2-D/3-D FPR
Before FPR, the average number of suspicious abnormalities was 291.6.
75
Sensitivity (%)
FPs/pass
After 2-D FPR After 3-D FPR
60 1.48 1.58
70 2.32 2.14
80 4.64 3.33
90 9.31 5.42
100 18.19 9.44
Reduce 94% Reduce 48%
FROC Curves
The FROC curves and the
corresponding JAFROC-1 figure of merit (FOM) for each feature set.
76
9.44 FPs/pass
Detected Results
A true positive case of 1.51 cm fibroadenomas
77
The potential tumor regions delineated
by watershed segmentation.
CAD The original ABUS image
Detected Results
A true positive case of 4.46 cm ductal carcinoma in situ
78
The potential tumor regions delineated
by watershed segmentation.
CAD
The dot circle indicates the false
positive.
The original ABUS image
Detected Results
A false positive case
79
The potential tumor regions delineated
by watershed segmentation.
CAD
The dot circle indicates the false
positive.
The original ABUS image
Discussion
We also applied the multi-scale Hessian analysis (published in IEEE TMI) to the same dataset used in this study to
provide a performance comparison.
– “Computer-Aided Tumor Detection Based on Multi-Scale Blob Detection Algorithm in Automated Breast Ultrasound Images,”
IEEE Transactions on Medical Imaging, vol. 32, no.7, pp.1191- 1200, 2013.
The proposed watershed method has better performance and short running time.
80
Detection rates 100% 90% 70% Running time
Hessian 18.0 9.1 4.3 13 minutes
Watershed 9.4 5.4 2.1 74.3 seconds
Summary
A fully automatic CADe system was proposed based on topographic
watershed for analyzing ABUS image.
The proposed CADe system showed sensitivities of 100% with 9.44
FPs/pass.
Rib and shadow regions were misclassified
– Further reduce the FPs.
81
Tumor Mapping
For ABUS, there are three views for a breast and a tumor will be
demonstrated in multiple views.
82
Experiment
A total of 53 abnormal passes with 41 biopsy-proven tumors and 13 normal passes were collected.
After CAD detection, a mapping pair was composed of a detected region in one pass and another region in another pass.
– Location criteria, including the radial position, relative distance and distance to nipple, were used to extract mapping pairs with close regions.
Quantitative intensity, morphology, texture and
location features were then combined in a classifier for further classification.
The performance of the classifier achieved a mapping rate of 80.39% (41/51), with an error rate of 5.97%
(4/67).
“Feasibility Testing: Three-dimensional Tumor Mapping in Different Orientations of 83
Automated Breast Ultrasound,” Ultrasound in Medicine and Biology, vol. 42, no.5, pp. 1201- 1210, 2016.
Mapping Case
84
L MED L AP
Benign fibroadenoma
ABUS/Mammo Mapping
85
Preliminary Clinical Result
Three readers: radiologist, surgeon, sonographer
– 168 ABUS views from 28 patients
– With CADe, more tumors could be found
Study design
– Step 1: Reader reviews the cases without CAD – Step 2: Reader reviews the missed CAD
markers at step 1
• Only the CAD marked missed at step 1 will be presented to reader.
• The missed CAD markers are decided by the distance of their nearest markers at step 1.
86
Step 1
Review without CAD
Step 2
Review the missed
CAD markers
Results
87
CADe Reader A Reader A
with CAD Reader B Reader B
with CAD Reader C Reader C with CAD
#TP 101 99 103 77 79 113 115
#FP 968 186 203 60 62 237 257
#FN 14 16 12 38 36 2 0
SEN 87.83 86.09 89.57 66.96 68.69 98.26 100
FP per
Pass 5.76 1.11 1.21 0.36 0.37 1.41 1.53
The missed CAD markers and their
BI-RADS scores were re-evaluated by reader at step 2.
Reviewing times of step 1 and step 2
88
Time per
View Step 1 Step 2
(CAD) Total CAD
Markers Time per Marker
Reader A 57.7 34.1 91.8 6.0 5.7
Reader B 26.0 12.7 38.7 6.1 2.1
Reader C 162.4 38.1 200.5 5.7 6.4
BI-RADS 3 4a 4b 4c 5 Total
Reader A 7 2 3 2 1 15
Reader B 2 1 0 0 1 4
Reader C 23 0 0 0 0 23
Case #1
Reader B
– Not marked at step 1 but marked at step 2 (CAD) – BI-RADS 4a
Readers A and C – Marked at step 1
89
Case #2
Reader B
– Not marked at step 1 but marked at step 2 (CAD) – BI-RADS 3
Readers A and C
– Not marked at step 1
90
Case #3
Reader B
– Not marked at step 1 but marked at step 2 (CAD) – BI-RADS 3
Readers A and C – Marked at step 1
91
Case #4
Readers A and B
– Not marked at step 1 but marked at step 2 (CAD) – BI-RADS 5
Reader C
– Marked at step 1
92
Free-hand Whole Breast US
Using magnetic tracker and image capturing, the conventional US can
be used to scan the whole breast, like ABUS.
93 trakSTAR
unit
Transmitte r
Frame Grabber Prob
e
Sensor
Magnetic Trackers
94
Transmitter Electronics unit
Sensor
Hitachi
GESensor
Free-hand WBUS System
The free-hand GPS system has been tried in the NTUH Breast Center and NTUH Yun-Lin Branch.
– The free-hand system will be compared with ABUS.
In the viewing system, the user can select any location to reviewing the images at that location.
95 Free-hand GPS Image Recorder Free-hand GPS Image Player
ABVS/MRI/TOMO Density Analysis
- RIB SHADOW FOR FINDING BREAST REGION
“Breast density analysis for whole breast ultrasound images”, Medical Physics, Vol. 36, No. 11, pp. 4933-4943, Nov. 2009.
“Comparative study of density analysis using automated whole breast ultrasound and MRI”, Medical Physics, Vol. 38, No. 1, pp. 382-389, Jan. 2011.
“Breast Density Analysis with Automated Whole-Breast Ultrasound: Comparison with 3-D Magnetic Resonance
Imaging,” Ultrasound in Med. & Biol., vol. 42, no.5, pp. 1211- 1220, 2016.
96
97
Aloka Whole Breast Density Analysis
Breast density measured from
mammograms and whole breast US
images Aloka whole breast US
Mammo
“Breast density analysis for whole breast ultrasound images”, Medical Physics, Vol. 36, No. 11, pp. 4933-4943, Nov. 2009.
98
Whole Breast US vs. Mammogram
Grade 3
Grade 4
Grade 2
U-systems Whole Breast Density Analysis
The similar Aloka whole breast
density analysis could be applied for the U-systems.
99
Original C View
Gland/Fat Analysis
U-Systems Density Analysis
The U-systems ABUS density and volume is highly correlated with those of MRI.
100
Density Volume
“Comparative study of density analysis using automated whole breast ultrasound and MRI”, Medical Physics, Vol. 38, No. 1, pp. 382-389, Jan. 2011.
MRI ABUS
FCM Density Analysis
The fuzzy c-mean (FCM) classifier was used to differentiate the fibroglandular and fatty tissues in ABUS and MRI images
101
Original
FCM
U-Systems Density Analysis
The U-systems ABUS density and volume is highly correlated with those of MRI.
102
Density Volume
“Comparative study of density analysis using automated whole breast ultrasound and MRI”, Medical Physics, Vol. 38, No. 1, pp. 382-389, Jan. 2011.
MRI ABUS
Rib Shadow for Finding Breast Region
For density analysis, an automatic breast segmentation method was
proposed based on the rib shadow to extract breast region from ABUS.
103
Transverse View
Sagittal View
An Example
Using rib shadow to extract the chest wall line.
104
Chest Wall Line
Segmented Breast Region
“Breast Density Analysis with Automated Whole-Breast Ultrasound: Comparison with 3-D Magnetic Resonance Imaging,” Ultrasound in Med. & Biol., vol. 42, no.5, pp. 1211-1220, 2016.
Experimental Results
MRI and ABVS images of 46 breasts from 23 women were collected.
Our results revealed a high correlation in WBV and BPD between MRI and ABVS.
Our study suggests that ABVS provides breast density information useful in the assessment of breast health.
105
TOMO Density
106
0.2012
0.3363 0.2110
MRI TOMO 2D Mammo
BI-RADS=3
From Dr. Moon, SNUH
LCC
TOMO Density
107
0.3685 0.2106 0.2318
MRI TOMO 2D Mammo
BI-RADS=4
From Dr. Moon, SNUH
LCC
Mammo Volumetric Density
MRI Density
MammoVolumetric Density
MRI CADe/CADx
“Computerized breast lesions detection using kinetic and morphologic analysis for dynamic contrast-enhanced MRI,”
Magnetic Resonance Imaging, vol. 32, no. 5, pp. 514-522, 2014.
“Computerized breast mass detection using multi-scale Hessian- based analysis for dynamic contrast-enhanced MRI,” Journal of Digital Imaging, vol. 27, no. 5, pp. 649-660, 2014.
“Computer-aided diagnosis of mass-like lesion in breast MRI:
Differential analysis of the 3-D morphology between benign and malignant tumors,” Computer Methods and Programs in
Biomedicine, vol. 112, no. 3, pp. 508-517, 2013.
“Computer-aided diagnosis of breast DCE-MRI using
pharmacokinetic model and 3-D morphology analysis,” Magnetic Resonance Imaging, vol. 32, no. 3, pp. 197-205, 2014.
109
Hessian-based Detection
Nonparametric nonuniform
intensity normalization algorithm (N4ITK) is used to reduce
nonuniformity bias.
The deformation of thorax to
approximate the outline between breast and chest by the Demons deformable algorithm is used to segment the breast region.
Fuzzy c-means algorithm (FCM) and Hessian algorithm are used to detect the tumors.
110
Bias field correction
Breast DCE-MRI
Motion correction
True tumor selection
Detected tumor Beast region segmentation
Tumor candidate detectionFCM
Hessian Preprocessing
Detection Result
111
Detection performance analysis
True tumor detection performance
– The mass detection rates are 100% (61/61)
with 15.15 false positives per case and 91.80%
(56/61) with 4.56 false positives per case.
112
Breast DCE-MRI CADx
The Pharmacokinetic Model and 3-D morphology Analysis are used to
diagnose the tumors in DCE-MRI.
113