Department of Information Technology

1
Computer Aided Staging of Prostate Cancer Using
MRI Image Processing Techniques Dept of
Computing and Information Technology, Philips
College
Funded by the Research Promotion Foundation of
Cyprus under code ?????/0104/03
2
Topics of Discussion

• Aim of proposed work
• Benefits
• Clinical Features
• MRI Imaging of Prostate Area
• Data Collection
• Overview of System Architecture
• Preliminary Results
• Summary

3
Aim of proposed work

• The aim of the proposed project is the design and
  the implementation of an automated system that
  will perform prostate cancer staging diagnosis.
• Staging of prostate cancer is performed mainly by
  radiologists and their decision is influenced by
  the findings on MRI images (radiological exams)
  of the prostate and the biopsy results (clinical
  results).
• Since most of the evidence (MRI) examination is
  performed with the naked eye, the proposed work
  is offering an automated way of classifying the
  several stages of the disease.
• Classification is achieved through the training
  of the proposed system with a number of cases
  that will result in specific recognition of
  patterns which will be used for the
  classification of the stages.

4
Benefits

• MR images are currently the best technique to
  assess problems associated with soft tissues
  (i.e. prostate).
• MRI uses a strong magnetic field and radio
  frequency waves to non-invasively obtain accurate
  morphologic images based on water tissue.
• Offers better localization of prostate cancer
  when combined with spectroscopy images
  (metabolic pictures based on the relative
  concentrations of cellular chemicals, i.e.
  choline, creatine, citrate).
• Therefore, using all existing MR imaging
  technologies (morphologic, dynamic and
  spectroscopy) for the accurate staging procedure
  can offer improved diagnosis and follow-up
  treatment

5
Clinical features

• Main concern is concentrated on
• Location and extent of tumor
• Extra-capsular extension of the tumor
• Seminal vesicle involvement
• Need to know the following clinical scores
• Prostatic Antigen Level (PSA) - blood test
• Gleason Score - biopsy
  
• PSA gt4ng/ml and Gleason gt4 are strong indications
  that the patient has a prostate tumor
• The exact location of the tumor as well as any
  possible extracapsular extension has an impact on
  the prognosis and diagnosis and follow-up
  treatment of the patient.

6
Staging categories TNM system

• Stage T1Tumor is microscopic, confined to
  prostate but undetectable by digital rectal exam
  (DRE) or ultrasound. Usually discovered by PSA
  tests or biopsies.Stage T2Tumor is confined to
  prostate and can be detected by DRE or
  ultrasound.
• T2a half of the lobe or less
• T2b more than one half of one lobe but not both
  lobes
• T2c both lobes Stage T3
• Tumor has spread to tissue adjacent to the
  prostate or to the seminal vesicles.
• T3a Unilateral extracapsular extension
• T3b Bilateral extracapsular extension
• T3c Tumor invades seminal vesicle(s)
• Stage T4
• Tumors have spread to organs near the prostate
  other than seminal vesicles

7
MR Imaging of Prostate CancerT2-weighted scans

• T2 images show the accurate anatomy and
  morphology of the prostate and peri-prostatic
  tissue
• Examples of axial images showing (a) peripheral
  and central gland (b) seminal vesicles

(b)
(a)
8
Examples of T2-weighted images
Seminal vesicle involvement
Suspicious areas on Peripheral zone (mid aspect
of both lobes
Both lobes on peripheral zone
9
T1- weighted scans

• This series of images better known as Dynamic
  Contrast Enhanced (DCE) MR images.
• These images show the permeability of blood
  vessels in the prostate area. This is achieved by
  injecting the patient with a contrast agent prior
  to scanning that starts enhancing with time.
  Rapid enhancement (high intensity) of suspicious
  areas is a strong indication of malignancy.
• The most common procedure involves the capture of
  eight axial images of the prostate, each of which
  is re-scanned every nine seconds.
• The result is a series of 240 T1-w images.
• The initial eight T1-w images are aligned and
  compared against their corresponding T2-w
  morphological images for better selection of the
  region of interest by the radiologist.

10
Example of a T1- weighted scan
11
Malignant vs. Normal Tissue
12
Metabolic Images

• Observe specific resonances (peaks) for citrate,
  choline and creatine from contiguous small
  volumes throughout the gland.
• The peaks for these different chemicals occur at
  distinct frequencies or positions in the MRSI
  spectrum
• The area under these peaks is related to the
  concentration of these metabolites, and changes
  in these concentrations can be used to identify
  cancer.

13
Data Collection

• Since October 2004, a number of prostate cancer
  cases have been collected for the purposes of
  this research project.
• The data was collected in the Radiological Center
  Ag. Therissos, where approx. 50 patients
  underwent MRI examination for prostate.
• MRI scanner
• Gyroscan NT Intera Philips Medical Systems at
  1.5 ?esla
• Coil SENSE-body
• The type of MR images that were collected were
• T2-weighted morphological images and
• T1-weighted DCE images.
• For the latter a Contrast Agent (Gadolinium DPA)
  was used for injection to patients prior to the
  scanning procedure.

14
Header file info of (MR) images
Since T2 and T1 images were of different size,
the two images had to be aligned accordingly so
as to have a pixel to pixel reference between T1
and T2.
15
Areas Under Investigation

• The proposed research project is deals with the
  following areas
• Noise removal of the MR images, alignment and
  motion correction.
• Feature extraction of T1-w and T2-w and metabolic
  image characteristics
• Fusion of extracted features with those resulting
  from spectroscopy images (Data fusion).
• Fusion algorithm that combines extracted features
  from the above types of MR images with clinical
  results, i.e. Gleason score, PSA.
• Pattern recognition and classification of stages.

16
Overview of System Architecture

• The proposed system comprises of 3 distinct
  sub-systems, namely
• (a) Data base system- responsible for the
  storage, search of data of the following type
• - MRI (T1, T2 and spectroscopy)
• - PSA levels
• - Gleason scores
• - biopsy reports (if MRI exam is post biopsy)
• (b) Graphical User Interface- it offers the
  following services
• Entry of new datasets to the database system
  (i.e. from optical disk, storage device)
• Selection of patient case and relevant data.
• Environment for the selection of region of
  interest (ROI)
• Visualisation of results

17
Overview of System Architecture (cont)

• (c) Sub-System for post-processing and automated
  staging- it involves the following sub-tasks
• Noise removal from ?1-W and ?2-W images,
  alignment and possible movement correction
• Extraction of dynamic features from T1-w images
  (permeability curves)
• Extraction of morphological features from T2-w
  images
• Feature extraction of spectroscopy images
• Fusion algorithm
• Detection and classification algorithm

18
(No Transcript)
19
Noise removal

• Noise on MR images is a result of the thermal
  noise produced by the receiver coils of the
  scanner as well as the electrolytes on the
  patient (Johnson noise). In our case where a body
  coil is used, the noise is greater since the
  scanned area is considerable (larger number of
  charged electrons or ions).
• It is widely accepted that this form of noise can
  be approximated as a random zero mean Gaussian
  noise. Therefore, in k-space (frequency domain)
  during the scanning process, the constructed
  image will be corrupted by additive Gaussian
  noise (complex).
• Since the final image is the amplitude of the
  Inverse Fourier Transform on k-space, it can be
  proved that the added noise on the image has a
  Rice distribution 1. More specifically
• Rayleigh disted on low intensity areas
• Normally disted on high intensity areas
• Aim reduce or eliminate the effect of noise but
  still preserve the edges of the image unaffected.
  
• Solution wavelets 1.
• 1 R. D. Nowak Wavelet Based Rician Noise
  Removal for Magnetic Resonance Imaging, IEEE
  Trans. on Image Processing, October 1999.

20
Motion correction

• Movement is produced on T1-weighted images due to
  either the patient movement or bowel movements.
• Therefore, an elastic motion estimation and
  correction method is required that will detect
  possible movements and correct them when
  necessary.
• Since the images involved are DCE images, this
  motion correction procedure should not affect the
  pixel intensity change that results from the
  contrast agent so as not to corrupt the resulting
  dynamic curves.
• Proposed Motion correction method
• Lucas Kanade2 minimize the sum of sq. error
  between a template image T and an image I warped
  back onto the coordinate frame of the template
• x-coordinate vector, W-affine warp, p- affine
  warp parameters
• In this case the derivatives of the images were
  used so as not to affect the image enhancement
  due to the contrast agent
• 2 B. Lucas, T Kanade. An iterative image
  registration technique with an application to
  stereo vision", Proceedings of the Int. Joint
  Conf on Artificial Intelligence, pp.674-679,
  1981.

21
Alignment

• The alignment technique employed, uses the fact
  that MRI images of the prostate area are
  symmetric (or approximately symmetric) with
  respect to a vertical axis.
• Why vertical axis
• invariant characteristic w.r.t. the
  particularities of the two different imaging
  modalities (T1 and T2 imaging).
• (due to improper patient placement during the
  scanning process, the symmetry axis might not be
  accurately vertical, but rather, approximately
  vertical).
• Alignment procedure involves
• estimation of the best vertical symmetry axes and
  angles (offsets from vertical direction),
  independently on each of the two images
• Exhaustive Search algorithm with optimization
  criterion the squared error between the left and
  right (flipped) sub-images with respect to that
  axis.
• rotation of images using interpolation (nearest
  neighbor, bicubic, bilinear).
• Having estimated the best possible vertical
  symmetry axis and angle, on both of the images, a
  scaling factor is estimated by examining the two
  images along their symmetry axes, determining the
  start and finish regions of both images, and
  dividing the corresponding lengths.
• Thus, a general affine transformation can be
  applied, that is invariant with respect to the
  particular characteristics of the two images.

22
Data Fusion

• The aim
• Combination of all extracted features,
• Dynamic Curves
• Morphological features
• Areas of high concentration ratios
• Gleason scores (if biopsy was performed prior to
  MRI exam)
• PSA
• The fusion procedure will offer reduction or
  elimination of errors that might occur in the
  scanning procedure (since it is unlikely that all
  scans will be systematically affected by the same
  errors).
• Moreover, the use of medical scores such as
  Gleason, or PSA will reinforce the validity of
  the results and their diagnostic value.

23
Detection and Classification

• The detection and classification algorithm
  constitutes the central part of the staging
  system.
• Detection deals with the estimation of a degree
  of in-homogeneity in the ROI under
  investigation.
• This in-homogeneity degree will reflect the
  difference on the behavior of several parameters
  in the ROI.
• High Degree suspicious
• Low Degree normal
• Classification uses Probabilistic Neural Networks
  (PNNs) 3 that incorporate all extracted
  features from all imaging modalities as well as
  the degree of in-homogeneity for use in the
  pattern recognition procedure.
• 3 D. F Specht, Probabilistic Neural Networks,
  Neural Networks, vol.3, 1990,pp.109-118.

24
Preliminary Results Classification

• The first step of the process involves the
  determination of normal and malignant
  lesions on the MRI images.
• Expert Radiologists marked several areas on the
  MRI images, as normal or Malignant.
• All patients used for the creation of the
  training data, had undergone a biopsy, that
  confirmed the malignancy of the areas marked
  under malignant.
• In order to help the Radiologist in marking an
  accurate Normal or Malignant area, the images
  used have been improved via digital processing(
  i.e. Wavelet Transform de-noising and motion
  correction)

25
Extracted Features

• The areas marked by the Radiologist, are
  transformed into a Training data-set, in such a
  way, that each pixel of each area corresponds to
  a separate feature-vector.
• Each feature-vector has a length of 35 features
  that are as follows
• Feature 1-30 The T1 dynamic curve that
  corresponds to that pixel
• Feature 31-33 T2 related characteristics
  (Intensity, local moments-not currently used)
• Feature 34-35 Factors related to how
  inhomogeneous the area of the vector is (not used
  for the current experiment)

26
Preliminary Results
27
Preliminary Results
28
Summary

• The proposed work deals with the design and the
  implementation of an automated system for
  prostate cancer staging.
• The system is based on the use of MR imaging
  techniques that is currently considered as the
  most appropriate for accurately describing the
  morphology of soft tissue, such as the prostate
  gland.
• Combination of morphological, contrast enhanced
  dynamic features and features that result from
  metabolic images through a data fusion scheme as
  well as the incorporation of clinical scores such
  as Gleason score or PSA can potentially result in
  a more accurate localization and estimation of
  the extend of prostate cancer.
• Preliminary results on classification based on
  the dynamic features of T1 images and the low
  signal intensity in T2 images, show the
  capability of PNNs to identify quite accurately
  the suspicious areas that have been independently
  identified by biopsy results.