Hyperspectral Imaging to Discern Benign and Malignant Canine Mammary tumors Control Sensor Network and Perception Laboratory Electrical and Computer Engineering.

Slides:



Advertisements
Similar presentations
Amrita Sahu 6th December, 2012
Advertisements

Hyperspectral two-photon near- infrared cancer imaging at depth Nikolay S. Makarov, Jean Starkey, Mikhail Drobizhev, Aleksander Rebane, Montana State University,
Quantifying soil carbon and nitrogen under different types of vegetation cover using near infrared-spectroscopy: a case study from India J. Dinakaran*and.
Trans-rectal near-infrared optical tomography reconstruction of a regressing experimental tumor in a canine prostate by using the prostate shape profile.
Lecture 10. The time-dependent transport equation Spatial photon gradient Photons scattered to direction ŝ' Absorbed photons Photons scattered into direction.
Alexander Vavilov Optical topography principles and example of an application. Bachelor Thesis Work.
Sergey Kucheryavski Raman spectroscopy Acquisition, preprocessing and analysis of spectra.
Face Recognition in Hyperspectral Images Z. Pan, G. Healey, M. Prasad and B. Tromberg University of California Published at IEEE Trans. on PAMI Vol 25,
R. Hui Photonics for bio-imaging and bio- sensing Rongqing Hui Dept. Electrical Engineering & Computer Science, The University of Kansas, Lawrence Kansas.
A NEW PERSPECTIVE TO VISIBLE NEAR INFRARED REFLECTANCE SPECTROSCOPY: A WAVELET APPROACH Yufeng Ge, Cristine L.S. Morgan, J. Alex Thomasson and Travis Waiser.
Soil Moisture Estimation Using Hyperspectral SWIR Imagery Poster Number IN43B-1184 D. Lewis, Institute for Technology Development, Building 1103, Suite.
Multiple Criteria for Evaluating Land Cover Classification Algorithms Summary of a paper by R.S. DeFries and Jonathan Cheung-Wai Chan April, 2000 Remote.
Hyperspectral Imagery
Rolando Raqueno, Advisor Credits for Winter Quarter, 2002: 2
Feature Screening Concept: A greedy feature selection method. Rank features and discard those whose ranking criterions are below the threshold. Problem:
Oral Defense by Sunny Tang 15 Aug 2003
Digital Technology 14.2 Data capture; Digital imaging using charge-coupled devices (CCDs)
SPECTRAL AND HYPERSPECTRAL INSPECTION OF BEEF AGEING STATE FERENC FIRTHA, ANITA JASPER, LÁSZLÓ FRIEDRICH Corvinus University of Budapest, Faculty of Food.
Introduction to Instrumental Analysis - Spectrophotometry
Introduction to Nuclear Medicine
Contrasting Precision Ag Technology Between Different Crop Species By Dodi Wear.
Integrated Fluorescent Probe and Radiofrequency Ablator Rachel Riti and Alex Walsh Advisers: Bart Masters and Anita Mahadevan-Jansen Department of Biomedical.
Remote Sensing Hyperspectral Remote Sensing. 1. Hyperspectral Remote Sensing ► Collects image data in many narrow contiguous spectral bands through the.
Spectral Characteristics
Miriam Israelowitz 1 and Dr. David L. Wilson 2 1 Department of Physics, Case Western Reserve University, Cleveland OH, 2 Deparment of Biomedical Engineering,
1 Exploiting Multisensor Spectral Data to Improve Crop Residue Cover Estimates for Management of Agricultural Water Quality Magda S. Galloza 1, Melba M.
Food Quality Evaluation Techniques Beyond the Visible Spectrum Murat Balaban Professor, and Chair of Food Process Engineering Chemical and Materials Engineering.
Photocapacitance measurements on GaP alloys for high efficiency solar cells Dan Hampton and Tim Gfroerer, Davidson College, Davidson, NC Mark Wanlass,
Pbio550: Biophysics of Ca2+ signaling ( washington
Metabolomics Metabolome Reflects the State of the Cell, Organ or Organism Change in the metabolome is a direct consequence of protein activity changes.
Michal Tepper Under the supervision of Prof. Israel Gannot.
Spectrophotometry.
Spectrophotometer.
Photoluminescence and Photocurrent in a Blue LED Ben Stroup & Timothy Gfroerer, Davidson College, Davidson, NC Yong Zhang, University of North Carolina.
Hyperspectral remote sensing
State of Engineering in Precision Agriculture, Boundaries and Limits for Agronomy.
Pixel Clustering and Hyperspectral Image Segmentation for Ocean Colour Remote Sensing Xuexing Zeng 1, Jinchang Ren 1, David Mckee 2 Samantha Lavender 3.
NIR.
Blackbox classifiers for preoperative discrimination between malignant and benign ovarian tumors C. Lu 1, T. Van Gestel 1, J. A. K. Suykens 1, S. Van Huffel.
Current 3D imaging systems for brain surgery are too slow to be effective in an operating room setting. All current effective methods for demarcation of.
Infrared IR Sensor Circuit Diagram and Working Principle.
DESCRIPTION OF PIXIRAD  The PIXIRAD Imaging Counter is an INFN-Pisa Spin-off  It works in photon counting  Pulse discrimination with two thresholds.
COMPARATIVE STUDY BETWEEN NEAR- INFRARED(NIR) SPECTROMETERS IN THE MEASUREMENT OF SUCROSE CONCENTRATION.
Electro-optical systems Sensor Resolution
Date of download: 6/9/2016 Copyright © 2016 SPIE. All rights reserved. Schematic showing the spatially modulated NIR illumination system. Figure Legend:
Date of download: 6/22/2016 Copyright © 2016 SPIE. All rights reserved. Schematic representation of the near-infrared (NIR) structured illumination instrument,
Date of download: 6/22/2016 Copyright © 2016 SPIE. All rights reserved. Photographs of exposed femoral bone surfaces and surrounding tissue prepared for.
Studies on the feasibility of using chemometric modeling of spectral data for the determination of post-mortem interval of skeletal remains. Kenneth W.
Date of download: 6/25/2016 Copyright © 2016 SPIE. All rights reserved. (a) Schematic of the interventional multispectral photoacoustic imaging system.
Date of download: 6/25/2016 Copyright © 2016 SPIE. All rights reserved. Lensfree imaging module. (a) Schematic illustrating the principle of lensfree image.
Date of download: 6/26/2016 Copyright © 2016 SPIE. All rights reserved. The absorption spectra of oxy- and deoxyhemoglobin relative to emission spectra.
Saturation Roi Levy. Motivation To show the deference between linear and non linear spectroscopy To understand how saturation spectroscopy is been applied.
In vivo monitoring of oxygen state and hemoglobin concentration
UV/VIS SPECTROSCOPY.
Optical Non-Invasive Approaches to Diagnosis of Skin Diseases
Mingyun Li & Kevin Lehmann Department of Chemistry and Physics
Chem. 133 – 3/14 Lecture.
In Search of the Optimal Set of Indicators when Classifying Histopathological Images Catalin Stoean University of Craiova, Romania
Bag-of-Visual-Words Based Feature Extraction
Computer Vision Lecture 4: Color
HyperSpectral Skin Imaging Tianchen Shi, Prof. Charles A. DiMarzio
OPTICAL MONITORING OF PHOTOSENSITIZER DIFFUSION INTO TISSUE
Hyperspectral Image preprocessing
What Is Spectral Imaging? An Introduction
Optical Non-Invasive Approaches to Diagnosis of Skin Diseases
An Infant Facial Expression Recognition System Based on Moment Feature Extraction C. Y. Fang, H. W. Lin, S. W. Chen Department of Computer Science and.
Somi Jacob and Christian Bach
Janis Spigulis, Vanesa Lukinsone, Martins Osis and Ilze Oshina
IT523 Digital Image Processing
Presentation transcript:

Hyperspectral Imaging to Discern Benign and Malignant Canine Mammary tumors Control Sensor Network and Perception Laboratory Electrical and Computer Engineering Department Temple University Philadelphia, PA 19122, U.S.A. Amrita Sahu 9 th May, Dr. Chang-Hee Won (Advisory Chair) Dr. Nancy Pleshko Dr. Joseph Picone 1

Outline Motivation Background and Literature Review Objectives Image Processing Methods Experimental Setup Characterization of System Data Acquisition Results Conclusions Future Work 2

Motivation Using Hyperspectral Imaging (HSI) for tumor detection Non-invasive Less time-consuming Allows assessment of a large area of tissue. Applications of HSI tissue imaging: Mammary tumors  Human Breast Cancer  Canine Cancer  Feline Cancer 3

Motivation No good device to identify malignant mammary tumors. Doctors usually perform biopsy or just ‘wait and watch’. Biopsy is the gold standard for cancer detection. It is invasive and requires several days for the results to be determined. To avoid the above disadvantages, we propose to use a non-invasive hyperspectral imaging sensor for characterizing canine mammary tumors. 4

Background Hyperspectral imaging measures and collects reflectance intensity information of more than hundred spectral bands across the electromagnetic spectrum. ton/journal/v3/n11/fig_tab/np hoton _F3.html 5

Applications of Hyperspectral Imaging Applications of Hyperspectral Imaging are: Agriculture Mineralogy Surveillance Monitoring of Oil Drilling Non-Invasive Tissue Analysis 6

Cancer Detection using Infrared Hyperspectral Imaging Breast Cancer Canine Mammary Cancer Tongue Cancer Gastric Cancer Skin Cancer Other diseases: intestinal ischemia, lung emphysema Liu, Z. et. al, “Tongue Tumor Detection in Medical Hyperspectral Images”, Sensors, 12(1), (2012). Balas, C., Themelis et. al “A Novel Hyper-Spectral Imaging System : Application on in-vivo Detection and Grading of Cervical Precancers and of Pigmented Skin Lesions”, In Proc. of "Computer Vision Beyond the Visible Spectrum" CVBVS'01 Workshop, Hawaii, USA, (2001). Lee, J., Won, C. H., “Characterization of Lung Tissues using Liquid-Crystal Tunable Filter and hyperspectral Imaging System,” Proc. IEEE EMBC 09, (2009). 7

Breast Cancer Results from 58 malignant breast tumors are reported. A steady state spectrometer used ( nm). Six laser diodes used for illumination. Fiber Optic cable delivers laser light to tissue. Shah, N., A. E. Cerussi, D. Jakubowski, D. Hsiang, J. Butler, and B. J. Tromberg, The role of diffuse optical spectroscopy in the clinical management of breast cancer. Dis. Markers 19:95–105,

Breast Cancer Hemoglobin, water and lipid content is different in malignant and benign tumors. Tissue Optical Index (TOI) was developed. Higher TOI indicates tumor malignancy. Shah, N., A. E. Cerussi, D. Jakubowski, D. Hsiang, J. Butler, and B. J. Tromberg, The role of diffuse optical spectroscopy in the clinical management of breast cancer. Dis. Markers 19:95–105,

Canine Mammary Cancer Fluorescent dyes used. The dyes were administered in the vein of the canine patient. For illumination, a 660nm laser diode beam used. The uptake and release rates of the dye varied in the diseased and normal tissue. M. Gurfinkel et al, Pharmacokinetics of ICG and HPPH-car for the Detection of Normal and Tumor Tissue Using Fluorescence, Near- Infrared Reflectance Imaging: A Case Study, Photochemistry and Photobiology, 2000, 72(1),

Objectives Characterize a hyperspectral imaging system and use it to discern malignant and benign canine mammary tumors. Normalize and preprocess the spectral data. Develop algorithms to discern malignant tumors and benign canine mammary tumors. Design an experiment to acquire the clinical data and analyze the results. 11

Why is Near Infrared Spectral Range Used? Near Infrared Hyperspectral Imaging has been used in literature for the detection of various kinds of cancer. NIR light has good penetration depth into tissue, because tissue has low absorptivity in this region. NIR light is absorbed by certain chromophores in the tissue that are biochemically significant. In this thesis, we use the visible - NIR spectral range, 650 nm to 1100 nm. 12

Methods 1. Image Preprocessing  To improve the signal to noise ratio.  Savitzky-Golay Smoothing process used.  It performs local polynomial regression using method of least squares. 2. Image Normalization  Data should be normalized to treat spectral non-uniformity of device.  Raw data may change due to illumination, temperature and non- uniform contour of the subject.  Range normalization used.  In range normalization, each spectral row is divided by its range (max value - min value). 13

Methods 3. Identification of chromophore-specific wavelengths  Second derivative method applied to the reflectance spectra.  Negative peaks in the second derivative spectra would give the wavelengths corresponding to the chromophores. 4. Algorithm to detect malignancy AlgorithmLiteratureCanine Cancer Support Vector MachineMost widely used in detection of prostate, gastric cancer. Does not work well PCA-LDAAlso used in some kinds of cancers. Sensitivity and Specificity 86 % and 86 % Tissue Optical Index (TOI)Used in breast cancerSensitivity and Specificity 86 % and 95% 14

Tissue Optical Indices Method A is the absorption of NIR light, ε is the molar extinction coefficient (mol/litre/cm) [c] is the concentration of chromophore (mol/litre) l is the photon path length (cm) 15

Tissue Optical Indices Method Higher content of hemoglobin (HbT) suggests elevated blood volume and angiogenesis. Higher water content (H 2 O) suggest edema and increased cellularity Decreased StO 2 (tissue oxygen saturation) indicated tissue hypoxia driven by metabolically active tumor cells Decreased lipid content reflects displacement of parenchymal adipose A higher TOI suggest that the tumor is malignant, because it indicates higher metabolic activity of the cells. 16

PCA-LDA Method Converts a larger number of correlated variables into a smaller number of linearly uncorrelated variables called principal components (PC). The first principal component has the highest variance, the second principal component has the second highest variance and so on. Each PC is orthogonal to each other. Principal Component Analysis / 17

PCA-LDA Method Linear Discriminant Analysis Linear Discriminant Analysis is widely used in statistics, machine learning and pattern recognition. It finds a linear combination of features which characterized two or more classes of objects. It used Bayes’ Formula, and we assume that the prior probabilities for groups are given. 18

PCA-LDA Method 19

Hyperspectral Imaging System Description The imaging system consists of: A digital imager (CCD camera, 1.4 megapixel, 12 bit output). A Liquid Crystal Tunable Filter. LCTF Controller. 500W dual quartz tungsten halogen lamps ( nm) were used for illumination. 20

Characterization of HSI System Experiment 1 To test the repeatability of the HSI system Experiment 2 Depth of penetration of NIR light into chicken breast tissue. Experiment 3 Effect of camera-to-sample distance on the reflectance intensity spectra. 21

Repeatability Experiment Hyperspectral image of a neoprene rubber sheet is captured for 5 consecutive days. The ambient temperature and humidity are recorded. External conditions such as lighting were kept as similar as possible. The reflectance spectra were compared. 22

Results DayTemperatureHumidity Weather conditions ˚C28%Cloudy ˚C29%Cloudy ˚C26%Sunny ˚C25%Sunny 524.1˚C25%Sunny System found to be fairly repeatable. 23

Depth of Penetration Experiment Chicken breast tissue was cut into sections of varying thickness. Neoprene rubber sheet was placed under the chicken slice. Quantify at what minimum width of the chicken slice the spectral effect of neoprene rubber is obtained. 24

Results The depth of penetration of NIR light into chicken breast to be between 3 mm to 5 mm 25

Effect of camera to sample distance on the output spectra. The distance between the sample and the camera is varied each time. The target was a slice of chicken breast. The reflectance spectra is not affected by the camera to sample distance. 26

Data Acquisition From Canine Patients The data were acquired in collaboration with the Veterinary Hospital of the University of Pennsylvania. 27

Data Acquisition From Canine Patients Hyperpspectral image cube of a canine patient Increasing wavelength 28

Results Spectral data smoothed by Savitzky-Golay filtering Smoothing applied on raw data to minimize noise 29

Results Smoothed spectral data from one of the canine patients shown. Cancer tissue has relatively lower reflectance intensity compared to the benign and the normal tissue. 30

Results After range normalization Before range normalization 31

Results Negative Peaks at 700, 840, 900 and 970 nm observed in the second derivative reflectance spectra, these peaks were attributed to deoxy- hemoglobin, oxy- hemoglobin, lipid and water respectively. Identifying chromophore-specific wavelength 32

Results (TOI Method)  Using a TOI threshold of 2.00 units, 6 out of 7 malignant tumors, 13 out of 15 benign tumors, All of 22 normal tissue ROIs were correctly identified.  Sensitivity and specificity of the proposed method were 86% and 95% respectively. 33 Sahu A., McGoverin C. et. al “Hyperspectral Imagimg to Discern Malignant and Benign Canine Mammary Tumors”, In Proc. of SPIE Defense Security Sensing 2013

Results (PCA-LDA Method) We have 44 Regions of Interest (ROI). Cannot construct separate training and testing dataset. 44-fold cross validation is used. Sensitivity and specificity is 86% and 86%. 34

Discussions Type of cancerSensitivity and specificity Prostate93% and 97% Gastric93% and 91% Skin90% and 75% Tongue93% and 91% Canine Mammary Cancer86% and 95% (TOI) 86% and 86% (PCA-LDA) Both TOI and PCA-LDA method works well. The TOI method has a slightly higher specificity for identifying benign tumors. 35

Dicussions TOI Method Advantages:  Four wavelengths identified characteristic of the four chromophores. This could significantly reduce imaging time.  Less-time consuming, easy to compute. Disadvantage:  Wavelength dependent PCA-LDA method Advantages:  More robust, takes into account all wavelength information.  Cross validation applied. Disadvantage:  More time consuming than TOI method. 36

Conclusions A hyperspectral imaging system was used to characterize malignant and benign canine mammary tumors. Reflectance intensities of malignant tumors were lower than benign and normal tissue over the wavelength 650 – 1100 nm. Four negative peaks were observed at the wavelengths of 700, 840, 900 and 970 nm characteristic of deoxyhemoglobin, oxyhemoglobin, lipid and water. A ‘Tissue Optical Index’ was used to classify canine cancer. Preliminary results with 22 canine mammary tumors showed that the sensitivity and specificity of the TOI method was 86% and 96% respectively. 37

Conclusions PCA-LDA method was developed to classify malignant tumors and the model was cross validated. The sensitivity and specificity of the PCA-LDA method was 86% and 86% respectively. 38

Future Work Further work needs to be done to collect more canine spectral data to generalize an application for the predictions put forward by the current study. We can also conduct an in-vitro study of the canine tumor tissue after resection and compare the analysis with that of the in-vivo study. In future experiments, we should use the reflectance standards to normalize data. 39

Future Work The lighting should be as uniform as possible. Non-uniform light will introduce variability in the data. The temperature of the tungsten halogen lights are very high. Uncomfortable for patients. Use of fiber optic cable can mitigate can the problem. 40

Acknowledgements I would like to thank the following people: Dr. Chang-Hee Won for providing me the opportunity to work on this project and for guiding me through the project. My Committee Members: Dr. Joseph Picone and Dr. Nancy Pleshko. Dr. Karin Sorenmo, for providing the canine patient data. Dr. Cushla McGoverin, for her constant help, support and encouragement. Dr. Won and Firdous Saleheen, for their help in canine data acquisition. The members of the CSNAP lab. Amrita Sahu is supported by the University Fellowship from Temple University Graduate School. This work was supported in part by the Tobacco Formula Fund of Pennsylvania Department of Health. 41

42