19
Computers in Biology and Medicine 36 (2006) 70 – 88 www.intl.elsevierhealth.com/journals/cobm An interactive interface for seizure focus localization using SPECT image analysis Mark Rossman a , Malek Adjouadi a , Melvin Ayala a , , Ilker Yaylali b a Department of Electrical & Computer Engineering, Florida International University, 10555W. Flagler Street, Miami, FL 33174, USA b Department of Neurology, Miami Children’s Hospital, 3100 S.W. 62ndAvenue, Miami, FL 33155, USA Received 20 May 2004; accepted 7 September 2004 Abstract Accurate epileptic focus localization using single photon emission computed tomography (SPECT) images has proven to be a challenging endeavor. First, commonly used radiopharmaceuticals such as hexamethylpropylene amine oxime (HMPAO) quantitatively underestimate large blood flows, leading to subtracted SPECT images that do not reflect the true cerebral physiological conditions, and often display non-distinct epileptic foci. The proposed relative change subtraction method of SPECT image analysis helps alleviate this quantitative burden. Second, the image analysis process traditionally performed by physicians is time consuming and prone to error. Toward this end, an automated algorithm was designed to analyze SPECT images and provide feedback to users through a visual interface. 2004 Elsevier Ltd. All rights reserved. Keywords: Epilepsy; HMPAO; Perfusion; SPECT subtraction; Medical images 1. Introduction In order to provide the physician with pertinent information before attempting epilepsy surgery, the imaging capabilities provided through the integration of single photon emission computed tomography (SPECT), magnetic resonance imaging (MRI), and electro-encephalograms (EEG) are often used to provide a reasonably accurate depiction of the in vivo physiology, allowing vascular malformations, Corresponding author. Tel.: +1 305 348 7054; fax: +1 305 348 3707. E-mail address: melvin.ayala@fiv.edu (M. Ayala). 0010-4825/$ - see front matter 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.compbiomed.2004.09.001

An interactive interface for seizure focus localization using SPECT image analysis

Embed Size (px)

Citation preview

Page 1: An interactive interface for seizure focus localization using SPECT image analysis

Computers in Biology and Medicine 36 (2006) 70–88www.intl.elsevierhealth.com/journals/cobm

An interactive interface for seizure focus localization usingSPECT image analysis

Mark Rossmana, Malek Adjouadia, Melvin Ayalaa,∗, Ilker Yaylalib

aDepartment of Electrical & Computer Engineering, Florida International University, 10555 W. Flagler Street,Miami, FL 33174, USA

bDepartment of Neurology, Miami Children’s Hospital, 3100 S.W. 62nd Avenue, Miami, FL 33155, USA

Received 20 May 2004; accepted 7 September 2004

Abstract

Accurate epileptic focus localization using single photon emission computed tomography (SPECT) images hasproven to be a challenging endeavor. First, commonly used radiopharmaceuticals such as hexamethylpropyleneamine oxime (HMPAO) quantitatively underestimate large blood flows, leading to subtracted SPECT images thatdo not reflect the true cerebral physiological conditions, and often display non-distinct epileptic foci. The proposedrelative change subtraction method of SPECT image analysis helps alleviate this quantitative burden. Second, theimage analysis process traditionally performed by physicians is time consuming and prone to error. Toward this end,an automated algorithm was designed to analyze SPECT images and provide feedback to users through a visualinterface.� 2004 Elsevier Ltd. All rights reserved.

Keywords: Epilepsy; HMPAO; Perfusion; SPECT subtraction; Medical images

1. Introduction

In order to provide the physician with pertinent information before attempting epilepsy surgery, theimaging capabilities provided through the integration of single photon emission computed tomography(SPECT), magnetic resonance imaging (MRI), and electro-encephalograms (EEG) are often used toprovide a reasonably accurate depiction of the in vivo physiology, allowing vascular malformations,

∗ Corresponding author. Tel.: +1 305 348 7054; fax: +1 305 348 3707.E-mail address: [email protected] (M. Ayala).

0010-4825/$ - see front matter � 2004 Elsevier Ltd. All rights reserved.doi:10.1016/j.compbiomed.2004.09.001

Page 2: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 71

lesions, and abnormal perfusion and electrical patterns to be visualized. Since SPECT demonstrates thefunctional activity (blood flow patterns) of the brain, it is considered especially valuable for diagnosingand localizing several neurological disorders, including epilepsy [1]. For this reason, the main focus ofthis research is placed on the detection and accurate localization of epileptic foci using ictal and interictalSPECT images.

Because SPECT only provides static images of an organ, multiple SPECT studies must be created inorder to analyze its dynamics. Many researchers have agreed that epileptic foci are reliably localized witha subtraction SPECT image [2,3]. This image is created via the information from two individual studies,where one study depicts the brain when it is not undergoing an epileptic seizure (interictal SPECT) andthe other depicts the brain during the epileptic event (ictal SPECT). This method is regarded as ‘basicsubtraction’ and is still widely used in the literature as the state of the art.

Areas of the brain that may be epileptogenic often display the following behaviors:

1. Between epileptic events, some brain tissue demonstrates that irritation exists through the dischargeof electrical spikes. At the same time, this irritation also results in a reduced blood flow in this tissue.

2. During seizure events, this tissue often experiences increased blood flow.

In this study, areas of the brain that satisfied criteria (1) and (2) above were cited as “probable” epilepticfoci by medical experts [4].

Subtraction between two SPECT scans is important because it detects the changes in cerebral perfusionof any pertinent regions across the image pair, thus highlighting their locations. Though SPECT imagesubtraction is effective at demonstrating changes in perfusion between studies, there are some constraintsto its utility. First, the process of obtaining subtracted SPECT images requires many processing steps andis interactive, requiring inputs from users familiar with brain dynamics. To date, programming tools forthe automation of SPECT image analysis are rarely mentioned in the literature. It would be advantageousif a system were made available that is fully automated. Second, the images obtained through SPECTsubtraction do not always depict a localized epileptic focus. Toward alleviating these burdens, an SPECTanalysis application was implemented in MATLAB that incorporates an improved SPECT subtractionmethod, relative change subtraction, and provides physicians with a turnkey solution with which toautomatically derive useful diagnostic information using only the images themselves as inputs.

2. Materials and methods

2.1. Patients

Ten patients with epilepsy were selected for this study by neurologists at Miami Children’s Hospital.Patients were selected that had recurring epileptic seizures. An intracranial EEG and both ictal andinterictal SPECT imaging studies were performed. A physician’s report detailing the most probableepileptic focus location, based on intracranial EEG, was provided for each patient. These assessmentsbased upon intracranial EEG were considered the “Gold Standard” by which the validity of the SPECTimage analysis algorithm was quantified. All of the procedures used in acquiring the necessary datafollowed strict protocols pursuant to the ethical guidelines and regulatory requirements regarding researchinvolving human subjects.

Page 3: An interactive interface for seizure focus localization using SPECT image analysis

72 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

2.2. Radiopharmaceutical, SPECT imaging system and acquisition parameters

The pharmaceutical hexamethylpropylene amine oxime (HMPAO) bound to Tc99m was used because ithas been shown to be effective in demonstrating cerebral blood perfusion [5,6]. Likewise, the efficacy ofHMPAO as an imaging tracer for epileptic studies is well accepted by the medical imaging community. Theadministered dose of Tc99m-HMPAO was compensated to a level between 5 mCi (185 MBq) and 20 mCi(740 MBq) using the patient weight (500 �Ci/kg). After the dosage was intravenously administered, eachpatient was then required to wait for a period of 90 min prior to the image acquisition procedure to allowfor proper uptake of the pharmaceutical.

A Siemens MULTISPECT3 triple-head gamma camera imaging system (Siemens, USA) was used toacquire the SPECT images. Each patient, after injection and waiting period, was scanned in a dimly-litroom with eyes closed, using a predefined protocol.

3. Automated system for SPECT image analysis

The contribution of this study was an automated SPECT image analysis algorithm, which included allof the steps shown in Fig. 1.

In the following subsections, a detailed description of the algorithm steps will be made.

3.1. Image registration

Registration is a necessary image pre-processing task whenever it is desired to compare two or moreimages that have been created from different physical viewpoints or using different imaging systems.Thus, for the purposes of this study, before meaningful subtractive SPECT images could be obtained,

Fig. 1. Automated SPECT processing algorithm.

Page 4: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 73

ictal and interictal images were co-registered to reduce the effects of patient motion between imagingsessions.

It is generally accepted that affine transformations can successfully correct for any study misalign-ments between intra-patient, intra-modality images [7,8]. Toward this end, an intensity-based registrationalgorithm was implemented that employs the affine transformation model with the optical flow error asthe similarity measure. The mathematical strategy to determine the optimal coefficients that correctlyregistered the image pair consisted of minimizing the optical flow error between the image pair usingspatio-temporal derivatives [9,10]. It was shown through a previous validation study that the registrationresults produced by the proposed algorithm were accurate enough for intra-patient SPECT brain stud-ies and gave comparable registration accuracies to those provided by professionally-produced softwarepackages [11].

3.2. Image intensity normalization

Knowledge of the radioactive distribution differences between imaging studies is important [12]. Sincesubtractive SPECT involves image pixel intensity differences, the values of these differences are partiallydependent on the amount of radioactivity in each study. To eliminate this activity difference, the studieswere first globally normalized [2]. For normal SPECT scans (acquired from patients without brain disor-ders), the ratio of mean intensities may be used as a metric for the activity differences between two images.However, for the case of SPECT scans of epileptic brains, other image intensity normalization metrics arenormally recommended. A method that has produced reliable results calculates the optimal normalizationfactor using the stochastic sign change (SSC) criterion [13]. Since it produced better accuracy, the SSCmethod of image normalization was used in this study.

3.3. Noise reduction

Artifact reduction was performed by using a simple thresholding technique. A thresholding exampleis shown in Fig. 2, where the reader can observe the influence of the applied percentage on the quality ofthe noise reduction.

Acceptable candidate threshold values can be found using a histogram analysis. The histograms shownin Figs. 3 and 4 are characterized by the presence of a local minimum. For 20 SPECT images, the averagenormalized intensity histogram has its major decline (inflection point) at approximately 5–10% intensityon a normalized scale. Through extensive research and visual verification, this threshold range was alsoshown experimentally to successfully reduce the SPECT image noise that was caused by the use of thefiltered backprojection method. After using an intensity threshold in the range of 5–10%, one will seethis artifactual noise disappear from the image and the smooth elliptical contour of the brain becomesapparent.

The SPECT image histogram can be subdivided into different regions as depicted in Fig. 5:

• 0–10%: noise (image reconstruction artifacts plus possibly Poisson noise);• 10–65%: SPECT brain pixels (not necessarily from epileptic lesions), moreover “background” brain;• > 65%: mix of non-epileptic brain pixels and lesion pixels (this applies to ictal SPECT images).

Page 5: An interactive interface for seizure focus localization using SPECT image analysis

74 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Fig. 2. SPECT image after thresholding.

Fig. 3. Normalized intensity histograms of 20 SPECT Images.

3.4. Image subtraction methods

To illustrate changes in cerebral perfusion between ictal and interictal brain states, the SPECT imageswere subtracted. Basic subtraction consists of applying the operation ‘ictal minus interictal’ to the two

Page 6: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 75

Fig. 4. Average normalized SPECT intensity histogram.

Fig. 5. A common intensity histogram for a SPECT ictal image, showing three well-defined regions of interest.

individual SPECT imaging studies. Even though basic subtraction is often effective, it sometimes lacksspecificity toward isolation of epileptic foci, even after proper image normalization has taken place. Tolessen this unwanted effect, an experimental subtraction method was proposed through research effortsinvolving the Miami Children’s Hospital Neuroscience Group. Percent change in perfusion, or relativechange subtraction, is the mathematical operation ‘basic subtraction divided by interictal’ as in Eq. (1). This

Page 7: An interactive interface for seizure focus localization using SPECT image analysis

76 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Fig. 6. Program outcome after applying the basic SPECT subtraction method to a patient.

augmented subtraction method is an attempt to normalize the subtractive SPECT intensities so that thepercentage change in blood perfusion between the studies can be observed. In previous studies, thismethod of subtraction demonstrated itself to have a better sensitivity and specificity for epileptic focuslocalization than basic subtraction [11]. This augmented the ability of the subtracted image to accuratelypinpoint the epileptic focus, increasing the utility of the image as a diagnostic tool.

Pixelresult[x, y, z] = Pixelictal[x, y, z]Pixelinterictal[x, y, z] − 1. (1)

Contrasting results of basic versus relative change subtraction are as shown in Figs. 6 and 7.For the patient case shown in Fig. 6, a very broad focus exists. It is C-shaped and extends from the

left temporal/parietal to the left frontal/temporal region. It is apparent that this case does not localizeeffectively to a small number of image slices with basic subtraction. The focus mentioned in Fig. 6 alsoappears in the relative change subtraction image in Fig. 7. However, in Fig. 7, the most intense portions(dark gray, black) only seem to concentrate in the central left temporal/parietal region in slices #34–#37.Thus, in this case, relative change subtraction improved the localization accuracy of the SPECT imagesto a fewer number of distinct slices: 4 using relative change subtraction (slices 34–37) versus 8–9 usingbasic subtraction (slices 29–37).

Due to the ability to produce better results than basic subtraction, only SPECT analysis algorithmpresented in this study.

Page 8: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 77

Fig. 7. Program outcome after applying the relative SPECT subtraction method to a patient.

3.5. Artifact reduction

One of the major challenges faced in this research work was to adequately remove artifacts that werefrequently leading to misclassification by the algorithm. A visual inspection of the percentage changesubtraction images of patients where those unwanted effects were detected revealed the presence ofoutliers. This was an unforeseen problem introduced by percent change subtraction. Outliers are shownas isolated voxels of higher intensity in a relatively low-intensity neighborhood, and are associated withnoise by medical experts.

Additional artifacts leading to misclassifications were observed in some patient data, where radioactivitywas absorbed by their eyes in amounts higher than expected. In those cases, image intensity in the eyeballregions was somewhat higher than the intensity in the lesion region, and therefore, the lesion was totallyundetected by the algorithm.

In an effort to remove these unwanted effects, artifact reduction in this study was performed in twosteps: firstly, removing the eyeball areas in the images, and secondly, eliminating the outlier pixels.

3.5.1. Eyeball detectionEyeball detection was performed to alleviate the burden of eyeballs being mis-detected as epileptic

foci. Medical experts state that, in some cases, the patient’s eyeballs can absorb some radioactivity afterinjection of SPECT radiotracer. Subsequently, the activity existing in the eyeballs will be represented inthe images as two equally-spaced bright regions in the upper two quadrants (regions 1 and 2, see Fig.15). Unfortunately, these bright regions often appear as false epileptic foci and, thus, may obscure thedetection of the true epileptic focus.

Page 9: An interactive interface for seizure focus localization using SPECT image analysis

78 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Fig. 8. Eyeball regions as detected by the algorithm on a subtracted image set. Rest of the image is set to background color ineach slice.

The implemented eyeball detection procedure employs a three-dimensional (3D) region growth (clus-tering) technique to detect symmetrically located pairs of regions of relatively high intensity in regions 1and 2 and selects the pair which best fulfills a set of constraints. An example of detection of eyeballs isshown in Fig. 8.

For a pair of pixel clusters existing in image regions 1 and 2, respectively, to be classified as eyeballs,they must satisfy the following constraints:

1. The average pixel intensity of each cluster must be greater than or equal to 35% of the subtractedimage maximum intensity.

2. Each cluster must be diametrically opposed about the mid-sagittal plane.3. Each cluster must occupy approximately the same set of coronal planes.4. Each cluster must occupy approximately the same set of transaxial planes.

After detecting the eyeball regions, their pixel values were set to negative. Since the algorithm does notperform any special processing on negative pixels, those areas have no influence in the results. Resultsof the application of this procedure for one patient are shown in Figs. 9 and 10.

3.5.2. Elimination of outliersFor percentage change subtraction images, the key to improving the epileptic focus detection seems to

lie in appropriate outlier reduction. Fig. 11 shows the image intensity histograms obtained after percentagechange subtraction. The peak observed around 7% can be explained by the average uptake of radiotracer

Page 10: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 79

Fig. 9. Subtracted images prior to eyeball detection.

Fig. 10. Subtracted images after eyeball detection.

Page 11: An interactive interface for seizure focus localization using SPECT image analysis

80 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Fig. 11. Intensity histogram after percentage change subtraction.

Fig. 12. A representative percentage change subtraction intensity histogram with intensity gap.

(HMPAO). It is commonly accepted by nuclear medicine community that the average cerebral uptake ofHMPAO is roughly 5–7% of the injected dosage in human patients.

A representative histogram is shown in Fig. 12, where several zero-crossing points can be observed.The intensity level represented in the first zero crossing point could be used as a first threshold for outlierremoval, in a hypothetical case where there are no lesion pixels above this intensity level. Empirical results

Page 12: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 81

show a relationship between the size of the intensity gap and the probability of encountering lesion pixelsabove this threshold. The higher the gap in the subtraction intensity histogram, the less probable it isto find lesions at intensity levels beyond this gap. Pixels far beyond this limit are more probable to beoutliers and less probable to be lesion pixels.

This interpretation is supported by the fact that the lesion should show a negative intensity gradient inthe interictal SPECT image as well as a positive gradient in the ictal image. Consequently, 3D subtractionshould contain enhanced gradients from brain background to the centroid of lesion. Big intensity gaps inthe resulting histogram cannot account for any gradient subtraction.

The procedure developed in this study to accurately remove outliers was based on voxel clusteringusing the zero-crossing intensity value for each image set. The steps are described as follows:

(1) detect starting threshold T of intensity gap on the subtracted 3D image;(2) perform voxel clustering on all voxels with intensity above T;(3) count voxels in each cluster;(4) for each cluster, apply following rule to classify cluster as outlier or lesion:

◦ classify the cluster: if number of voxels > 8, then cluster is a lesion cluster, otherwise, it is anoutlier cluster;

◦ if cluster is an outlier cluster, then remove cluster by setting all voxel values to zero;

(5) set highest voxel intensity to max.

The significance of the number 8 mentioned above can be explained as follows: a classification criterionwas proposed [4] that stated that a true lesion should occupy an image volume of at least 1 cm3; thiseliminated small groups of pixels from the image as noise. In terms of the SPECT imaging system usedin this study, which had an isotropic voxel resolution of 3.56 mm, this suggested that a minimally-sizedlesion would occupy approximately a space of 3 × 3 × 3 voxels. This is equivalent to assuming that thelesion is wider than 2 pixels in each dimension, assuming a cubic-shaped lesion for simplicity. Thus, alesion could be assumed to occupy greater than 8 voxels in image space.

3.6. Image intensity scaling and color display

Grayscale display techniques are usually employed in the traditional SPECT image interpretationprocess. Since the human eye can only differentiate a relatively small number of gray levels, the contrastof these images may appear limited [14]. A proposition was to display images with pixel intensities madeproportional to a color map, as shown in Fig. 13. This type of display is not only more appealing butalso helps in the quantitative image interpretation process. The main window of the application is shownin Fig. 14. Interaction with the main window will activate additional windows that show the subtractedSPECT images in full color.

3.7. Epileptic focus localization method

Once a subtracted SPECT image was created using the aforementioned steps, determining the locationof the epileptic focus remained. An epileptic focus is generally defined as a concentrated group of image

Page 13: An interactive interface for seizure focus localization using SPECT image analysis

82 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Fig. 13. SPECT multi-color display application.

Fig. 14. Main window of the application. This window is used for file opening and options selection. It also displays the namesof the proposed regions along with additional windows that show the processed images.

pixels whose values have a similar high intensity, which is analogous to high changes in perfusion.Generally, thresholding can be used to separate foci pixels from background pixels.

Once a focus is isolated through thresholding, it needs to be localized and assigned a location in thebrain. This technique used nomenclature that was consistent with the major regions of the brain (frontal,temporal, parietal, or occipital lobes). A simplified 2D model of the brain was thus proposed, as illustratedin Fig. 15.

The above representation of the brain was created to subdivide it into six major anatomical regions asfollows: region 1: right frontal lobe, region 2: left frontal lobe, region 3: right temporal lobe, region 4:left temporal lobe, region 5: right occipital/parietal lobe, region 6: left occipital/parietal lobe.

Page 14: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 83

Fig. 15. Simplified 2D brain model.

Fig. 16. Two SPECT Images after application of steps 1–4 of the localization procedure.

Focus localization in 3D subtracted SPECT images involves the following tasks:

1. Using the pixels of the mid-axial 2D slice image, calculate the values Xmid, Ylower, Yupper.2. Threshold the 3D subtracted SPECT image.3. Find the coordinates (x, y, z) of all remaining non-zero image voxels.4. Using the (x, y) coordinate of each voxel and the values of Xmid, Ylower, and Yupper, determine to

which region (1–6) it belongs.5. Count the total number of voxels in each region.

This set of tasks produced results similar to those shown in Fig. 16 and Table 1. For each thresholdvalue, a pixel count of the image by region was performed. Based on these pixel counts, a determinationabout which region(s) had the highest pixel populations was made.

As the algorithm completed its tasks, it produces a description of the location of the seizure activity.Since the localization result is dependent on the value of image intensity threshold used, this value is also

Page 15: An interactive interface for seizure focus localization using SPECT image analysis

84 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Table 1Pixel count of SPECT images in Fig. 8

Region Case 1 Case 2

# pixels in region 1 48 9# pixels in region 2 6 11# pixels in region 3 14 63# pixels in region 4 8 58# pixels in region 5 5 7# pixels in region 6 5 13Total # of pixels 86 161Overall focus region Region 1 Regions 3, 4

Table 2Seizure focus localizations for the case shown in Figs. 3 and 4 (�: threshold mean, �: threshold standard deviation)

Patient No. 2 Basic subtraction Relative change subtraction

Range 1 4 4Range 2 4 4Range 3 4 4Range 4 4, 6 4� + 1� 4, 6 4� + 1.5� 4 4� + 2� 4 4� + 2.5� 4 4Vote per region 8 for reg. 4 8 for reg. 4

2 for reg. 6Vote count 10 8Vote percentage 80% for reg. 4 100% for reg. 4

20% for reg. 6

included in the results. Shown in Table 2 is an example of this focus localization procedure using one setof patient SPECT data. For this case, the automated algorithm predicted Region 4 (left temporal lobe) asthe brain region with the highest probability of containing the epileptic focus.

In Table 2, threshold range 1 is 50% for basic subtraction and 5% for relative change subtraction,respectively. For ranges 2–4 the values are 60% and 10%, 70% and 15%, and 80% and 20%, respectively.The last percentage values (5%, 10%, 15% and 20%) were found to yield the most significant changes inthe overall results [11]; therefore, these values were the ones used for relative change subtraction in thisstudy.

4. Algorithm performance analysis

4.1. Receiver operating characteristic

The localization results provided by the algorithm were gauged against expert medical assessmentsfor each case using analysis. This analysis gave a quantitative result as to the sensitivity, specificity, andaccuracy of the proposed algorithm.

Page 16: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 85

Table 3Sensitivity, specificity and accuracy of automated SPECT analysis as a function of threshold value (�: threshold mean, �:threshold standard deviation)

Thresh. TPF TNF Accuracy(sensitivity) (specificity)

5% 0.67 0.93 0.8910% 0.83 0.97 0.9415% 0.83 0.97 0.9420% 0.83 0.97 0.94� + 1� 0.50 0.90 0.83� + 1.5� 0.83 0.97 0.94� + 2� 0.67 0.97 0.91� + 2.5� 0.67 0.97 0.91

For each patient, the six regions of the brain existed as separate entities to be tested for the presence orabsence of an epileptic focus. Since there were ten patients, this gave 60 separate entities to be tested. Thealgorithm gave the following answers: 0 (indeterminate region), 1–6, or any combination of two adjacentbrain regions (1,2; 1,3; 2,4; etc.). Likewise, the medical experts’ assessment consisted of a single brainregions (regions 1–5 or 6) that contained the epileptic focus. As stated earlier, the physicians’assessmentswere derived from the visual inspection of intracranial EEG obtained from each patient. The ten individualbrain regions (1 focus region for each patient) were the only “positive” or “abnormal” brain regions; theremaining 50 individual brain regions were automatically declared as “negative” regions. Based on thistesting scenario, the sensitivity, specificity, and accuracy of the automated SPECT analysis algorithm werecalculated for different threshold values (Table 3). Optimal performances were obtained with thresholdsequal or greater than 10% of the normalized intensity scale or beyond the measure computed as �+ 1.5�.

Such positive outcomes could be integrated with EEG- or MRI-based analysis for added validation[15,16].

It can be observed from Table 3 that the highest parameters corresponded to the thresholds 10%, 15%,20%, as well as the mean + 1.5 SD threshold.

The aforementioned range of thresholds was used in the form of a committee (as shown in Table 4) toeliminate the dependence of the system performance parameters on a single threshold value. Specifically,this study uses all the thresholds shown in Table 3 in a committee and attempts to find a solution byusing the contribution of each one of the thresholds to create a composite result, so as to consolidate theregion-based agreement. Results of this process are shown in Table 4. The performance of the systemlooks poor for patient #7, because the epileptogenic zone spreads to neighboring region.

In Table 4, the 2nd column represents the regions declared by the physicians to contain the epilepticfocus, whereas the following eight columns show the predominant region, which was determined usingeach one of the individual thresholds. The algorithm then summarizes the results (last column) by countingthe number of occurrences of each detected region and dividing it by the total number of possible detections(8). This percentage of agreement between the threshold-based committee members is the final output ofthe proposed algorithm.

Page 17: An interactive interface for seizure focus localization using SPECT image analysis

86 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Table 4Physician assessments and final results obtained after running the system with available data

Patient # Phys. focus 5% 10% 15% 20% �+1� � + 1.5� � + 2� � + 2.5� Agreement

1 3 3 3 3 3 3 3 3 3 100% (3)2 3 3 3 3 3 3 3 3 0 87.5% (3)

12.5% (0)3 3 3 3 3 1 3 3 1 1 62.5% (3)

37.5% (1)4 3 3 3 3 0 3 3 3 3 87.5% (3)

12.5% (0)5 2 2 2 2 2 2 2 2 2 100% (2)6 4 4 4 4 2 4 4 4 2 75% (4)

25% (2)7 4 4 4 2 2 4 4 2 2 50% (4)

50% (2)8 1 1 1 1 0 1 1 1 1 87.5% (1)

12.5% (0)9 2 2 2 2 0 2 2 2 0 75% (2)

25% (0)10 1 1 1 3 3 1 1 1 3 62.5% (1)

37.5% (3)

Table 5Time distribution of the computational steps needed for each individual patient

Pat. no. Total execution time (s) Pat. no. Total execution time (s)

1 103.71 6 ()105.682 106.69 7 ()135.943 132.10 8 131.614 98.33 9 ()133.255 103.07 10 ()148.48

4.2. System timing analysis

A key factor in automated processes is the time required for finishing all the computational steps. Thisapplication underwent a detailed timing analysis based on the patients under study. Table 5 shows thetime needed for the algorithm to process each set of SPECT data.

The average time for execution was approximately 119 s (2 min) using a Pentium IV, 3 GHz processorwith 256 MB RAM, which is an acceptable time for this type of off-line analysis. These observationsdemonstrate that the developed system is suitable to be used in clinical environments due to its relativelyshort execution time.

5. Conclusion

The overall benefits of this research were realized in that an automated program was engineered thatquickly and accurately performs a set of integrated steps normally tedious and otherwise time consuming

Page 18: An interactive interface for seizure focus localization using SPECT image analysis

M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88 87

and requiring a human operator who is skilled in brain anatomy. Therefore, this program is a directbenefit to physicians, as it shortens the time necessary to assess each patient case, allowing more patientsthe opportunity to receive medical assistance. Additionally, it was determined that, for the ten sets ofpatient SPECT image data, the automated SPECT image analysis algorithm obtained an impressive91.3% average rate of epileptic focus localization accuracy. Therefore, it has been demonstrated that thisalgorithm gives reliable results and can be used as a physician aid for epilepsy cases.

Based on these encouraging results, the authors believe they have made a modest contribution to theautomation of the processing and interpretation of SPECT images.

Acknowledgements

The authors wish to thank the National Science Foundation and the Office of Naval Research for theirsupport provided under Grants EIA-9906600, HRD-0317692 and N-000149910952.

References

[1] A.S. Harvey, Temporal lobe epilepsy in childhood, Ph.D. Thesis, University of Melbourne, 1993.[2] B.H. Brinkmann, T.J. O’Brien, J. Terence, B.P. Mullan, M.K. O’Connor, R.A. Robb, E.L. So, Subtraction ictal SPECT

coregistered to MRI for seizure focus localization in partial epilepsy, Mayo Clin. Proc. 75 (6) (2000) 615–624.[3] H.W. Lee, S.B. Hong, W.S. Tae, Opposite ictal perfusion patterns of subtracted SPECT, Brain 123 (2000) 2150–2159.[4] R.A. Avery, I.G. Zubal, R. Stokking, C. Studholme, M. Corsi, J.P. Seibyl, S.S. Spencer, Decreased cerebral blood flow

during seizures with ictal SPECT injections, Epilepsy Res. 40 (2000) 53–61.[5] J.K. Moretti, M. Caglar, P. Weinmann, Cerebral perfusion imaging tracers for SPECT: which one to choose?, J. Nucl. Med.

36 (1995) 359–363.[6] S. Asenbaum, T. Brucke, W. Pirker, U. Pietrzyk, I. Podreka, Imaging of cerebral blood flow with technicium-99m-HMPAO

and technicium-99m-ECD: a comparison, J. Nucl. Med. 39 (1998) 613–618.[7] R.P. Woods, Within-modality registration using intensity-based cost functions, in: I. Bankman (Ed.), Handbook of Medical

Imaging: Processing and Analysis, vol. 33, Academic Press, New York, 2000.[8] J.B.A. Maintz, M.A. Viergever, A survey of medical image registration, Med. Image Anal. 2 (1) (1998) 1–37.[9] S. Mann, R.W. Picard, Video orbits of the projective group: a simple approach to featureless estimation of parameters,

IEEE Trans. Image Process. 6 (9) (1997) 1281–1295.[10] M.A. Rossman, F. Candocia, M.Adjouadi, P. Jayakar, I.Yaylali,Application of affine transformations for the co-registration

of SPECT images, Proceedings of the Fourth IASTED International Conference on Signal and Image Processing, August2002, pp. 595–600.

[11] M. Rossman, M. Adjouadi, N. Mirkovic, M. Ayala, P. Jayakar, I. Yaylali, An integrated approach to localize epileptic fociusing relative SPECT subtraction, Proceedings of the IASTED International Conference on Modeling and Simulation,ISBN: 0-88986-337-7, Palm Springs, CA, USA, February 24–26, 2003, pp. 342–347.

[12] A.F.G. Rocha, J.C. Harbert, Textbook of Nuclear Medicine: Basic Science, Lea and Febiger, Philadelphia, 1978.[13] J.B.A. Maintz, Retrospective registration of tomographic brain images, Ph.D. Thesis, University of Utrecht, 1996.[14] B.K.P. Horn, Robot Vision, McGraw-Hill, New York, 1986, pp. 278–293.[15] N. Mirkovic, M. Adjouadi, I. Yaylali, P. Jayakar, 3-D source localization of epileptic interictal spikes, Brain Topography

16 (2) (2003) 111–119.[16] M. Adjouadi, D. Sanchez, M. Cabrerizo, M. Ayala, P. Jayakar, I. Yaylali, A. Barreto, Interictal spike detection using the

Walsh transform, IEEE Trans. Biomed. Eng. 51 (5) (2004) 868–872.

Mark Rossman received his Bachelor of Science in Electrical Engineering in 1999 and his Masters Degree in ElectricalEngineering in 2003 from FIU. His research interests include analog electronics, electromagnetism, and digital image processing.

Page 19: An interactive interface for seizure focus localization using SPECT image analysis

88 M. Rossman et al. / Computers in Biology and Medicine 36 (2006) 70–88

Currently he is pursuing his Master of Science in Computer Engineering. He is working in conjunction with Miami Children’sHospital on thesis research focused on medical image processing.

Malek Adjouadi obtained his B.S. in Electrical Engineering from Oklahoma State Univ. in 1978, his M.E. and Ph.D. degreesboth from the Univ. of Florida in 1981 and 1985, respectively. Dr. Adjouadi is currently serving as Associate Professor andis founder and Director of the Center for Advanced Technology and Education established by NSF and ONR grants at theElectrical & Computer Engineering Department from Florida International University. His interests include computer vision,image processing, human computer interfaces, and applications of Neuroscience.

Melvin Ayala obtained his Bachelor in Industrial Engineering in 1984 and his Ph.D. in 1987 at the Zittau Engineering Institute,Germany. He has served as a professor and researcher in Cuba and Brazil. Dr. Ayala is currently working as a Research Associatewith the Electrical & Computer Engineering Department at Florida International University, and is currently serving as managerat the Center for Advanced Technology and Education. His fields of interest are in software development, pattern recognition,image/signal processing, artificial neural networks and fuzzy logic.