& Over 50 Breast

• Geometrical features based method.

• Statistical features based method

• 3Db features based method.

C A

D - C T C P o l y p D e t e c t i o n

C T C olonography is a rapidly evolving technology for th e detection of colorectal polyps and m any studies have d em o n strated th a t its sensitivity in polyp detection is com parable to th e sensitivity offered by conventional colonoscopy [21, 23, 26, 29, 32, 33, 34], In th is regard, Fenlon et al. [21] indicate th a t C T C retu rn s 100% sensitivity for th e detection of C T C polyps greater th a n

10

m m and 83% sensitivity for detectio n of polyps in th e range 6-9mm polyps. T his conclusion is su pp orted by oth er studies [32, 74, 75, 76] w here it is d em onstrated th a t C T C is as good as sta n d a rd colonoscopy for th e detection of colonic polyps. M ore recently, P ick h ard t e t al. [29] perform ed a detailed com parison between C T C and stan d ard Colonoscopy and th e y concluded th a t C T C can increase th e sensitivity of polyp detection w hen applied as a second reader w ith Colonoscopy. In th eir study, th e rep o rted sensitivities for C T C and optical colonoscopy for polyps > 10m m were 92.2% and 88.2%, for polyps > 9m m were 91.8% an d 90.2%, for polyps >

8

m m were 92.6% and 89.5%, for polyps > 7m m were 89.5% and 90.2%, for polyps >

6

m m were 85.7% and 90.0% respectively. From these results th ey concluded th a t th e sensitivity in polyp d etection offered by C T C m atches closely th e sensitivity achieved by optical colonoscopy and C T C is feasible to be used in clinical exam inations.

Since th e in tro d u ctio n of C T C in 1994 [15], a large num ber of techniques in th e fields of 3D visualization, such as th e rendering of th e colon surface, centerline calculation, and colon wall unfolding were developed to provide th e radiologists w ith all types of 2D an d 3D inform ation required to identify th e colorectal polyps. [77, 78, 79, 80, 81, 82, 83, 84, 85,

86

, 87]. However th e developm ent of new C T im aging m odalities, th e high resolution C T d a ta offers a large volume of inform ation th a t is required to be visualized and in terp reted by th e radiologists (the typical tim e required to process a d atase t based on a visual exam ination is in th e range 12-60

m inutes). As pointed out in th e stu d y by P ick h ard t et al. [29] th e perform ance of th e radiologists can be effected by factors such as p ercep tu al errors [

88

, 89] and eye fatigue [34, 35]. Johnson et al. [

88

] stu d y shows th a t 34% (20 of 59) of th e large polyps were m issed in C T C due to p ercep tu al errors. Hence, Ven Gelder et al. [89] suggested th a t th e in tro d u ctio n of CAD based au to m atic polyp detection in C T C is a viable solution to reduce th e p ercep tu al errors associated w ith th e visual in terp reta tio n of th e C T C d atasets. T hus, th e developm ent of C AD m ethods can im prove b o th th e sensitivity and efficiency of CTC. In th e last decade a significant am ount of research has been focused on developing au to m ated CAD of colonic polyps and a large num ber of C A D -based polyp detection techniques have been proposed.

One of th e first C A D - C T C system s was proposed by V ining et al. [90] w here th e detection of colonic polyps was based on surface cu rvature analysis. In th e exper­ im ental section of th e ir p ap er th ey indicated th a t th e C A D -C T C system achieved 73% sensitivity w ith 9 to 90 false positives (F P )/d a ta s e t.

T he polyp d etectio n system developed by Sum m ers et al. [91] atte m p ts to iden­ tify th e polyps in th e C T d a ta using a m ulti-stage geom etrically-driven approach. Initially, th e y detect th e convex surfaces th a t p ro tru d e inw ard from th e colon by ap­ plying a kernel filer th a t is con structed using p a rtia l derivatives. A fter th e detection of th e cand id ate surface, th e y used shape-based criteria derived from th e principle curv ature (kmin an d kmax), m ean curvature (H), sphericity ratio s = {kmax — kmin) / H and m inim um polyp size. T h ey used very restrictive sphericity criteria in order to reduce th e false positives b u t th eir technique shows zero sensitivity for polyps in th e range 5-10mm (0 o u t of 4) and 100% sensitivity for polyps > lOmm

(6

out of

6

). L ater, Sum m ers et al. [36] proposed a new m eth o d m eth od where th ey applied a different shape based filter (calculated from kmin, kmax, and H ) to reduce th e level of F P b u t keeping th e sensitivity a t 100%. One problem w ith th is approach is th e fact th a t th e sensitivity and specificity of th e system depend on th e filter chosen to evaluate th e local colon cu rvature and th e rep o rted sensitivities in polyp detection are in th e range 29% to 100% w ith

6

to 20 F P s /d a ta s e t.

Yoshida et al. [37, 92] proposed th e use of shape index (cup, ru t, saddle, ridge, cap), curvedness values (calculated on small volumes of interest) and fuzzy clustering in order to perform can d id ate polyp surface generation. T he principal curvature {kmin and kmax) derived from th e G aussian and th e m ean curvature was used to calculate th e shape index and curvedness for each colonic wall voxel. They showed

th a t all types of colon shapes can be m apped in th e interval S 7 e[0 ,1] as follows: cup (0.0), saddle (0.5), ridge (0.75), and cap (1.0). O n th e o th er hand, th ey showed th a t th e curvedness is also an indicator of th e variation of th e local curvature and they used a predefined threshold w ith a value betw een 0.9 to 1.0 for SI and 0.08m -1 to 0.20m m

-1

for curvedness to generate th e initial seed points. T he C-M eans clustering was used to generate th e cand id ate surfaces and to reduce th e incidence of non polyp surface generated by noise. T he C A D -C T C system [92, 37] employed features such as th e shape index, curvedness, m agnitude of C T values, C T values, gradient concentration (GC) a n d direction of th e gradient concentration (DGC) calculated from candidate surfaces to classify th em into polyps or folds. T hey repo rted 95% sensitivity in polyp detection w ith 1.2 F P per d atase t, b u t th e F P s increased w ith a factor of 1.5 w hen th e sensitivity was increased to 100%.

P aik et al [35, 93] developed a new algorithm called surface norm al overlap th a t was applied to colorectal polyp detection. T heir algorithm is based on th e assum ption th a t th e colorectal polyps are convex stru ctu res and th e local norm al intersection density sam ples th e local convexity for each voxel of th e colon wall. The norm al overlap technique was used to identify suspicious convex stru ctu res while th e polyp detection is perform ed by assessing th e deviation of these convex stru ctu res from a stochastic m odel employed to define th e shape of a nom inal polyp. This algorithm shows 100% sensitivity in detecting polyps larger th a n 10mm w ith 7 F P datasets. No experim ental d a ta is provided in regard to th e sensitivity of their CAD- C T C system w hen applied to th e identification of sm all (< 5m m ) and mid-sized polyps (betw een 5-10mm).

Kiss et al. [94, 95, 96] m eth od also employed th e surface norm al intersection for th e detection of convex surface from th e colonic wall. To generate th e polyp cand id ate surface, th e y applied th e Hough Transform to calculate th e center points an d used 3D region growing to find th e cand id ate surface from th e convex voxels of th e colon wall. G aussian d istrib u tio n of th e H ough points was used to calculate th e norm al concentration of th e candidate surface. Two different region growing techniques (weighted region growing and greedy region growing) were employed to generate can didate surfaces from th e center points and least square ellipsoid fitting was used to calculate th e th ree axes of th e can didate surface resulting from these two region growing algorithm s. T he num ber of norm al intersections for each Hough point, G aussian d istrib u tio n , th ree axes of th e surface resulting from th e greedy

region growing and th re e axes of th e surface generated by the weighted region grow­ ing were used as in p u t features for a probabilistic n eural netw ork (PNN ) classifier. T h eir C A D -C T C system achieved 90% sensitivity for polyps larger th a n

6

mm w ith 2.82 F P s /d a ta s e t. Recently, a different C A D - C T C system has been proposed by Kiss et al. [38] th a t analyses th e slope density function as a discrim inative feature to classify th e convex can d id ate surfaces into polyps and folds. T he initial stage of th e ir system identifies th e can d id ate surfaces by intersecting th e colon wall with a refo rm atted plane p erp en d icular on th e local norm al surface. If th e intersection patch between th e p la n ar and th e colon surface is filled w ith voxel d a ta th e colon surface is concave an d is declared p a rt of th e h ealth y colon tissue. O therw ise is a convex surface th a t is generated either by polyps or folds. T he resulting can­ d id a te surfaces are evaluated statistically using th e slope density function, which shows peaks for elongated surfaces an d sm ooth values for ellipsoidal surfaces. This p ro p erty of th e SDF is very useful as it provides ro b ust discrim ination betw een th e polyps and folds as th e folds resem ble elongated cylindrical surfaces w hereas polyps ellipsoidal surfaces. T h eir m eth od obtained th e following perform ance in polyp de­ tection: 33.33% sensitivity for polyps sm aller th a n 5mm, 85.70% for polyps in th e range 6-9mm, 90% for polyps larger th a n 9mm and 100% sensitivity for cancerous lesions. K iraly et al. [97] proposed a fast d etection m eth od using a gradient-based filter and shows 96% sensitivity for polyp greater th a n 5mm w ith 5.76 false positive per d ataset.

A car et al [39, 98] employed a different approach based on th e edge flow displace­ m en t th a t is applied to ob tain ro b u st polyp detection. T hey developed a m ethod to ex tra ct th e can d id ate surfaces based on th e H ough Transform th a t evaluates th e no rm al intersections using th e assum ption th a t th e norm al intersection will be high for convex (cap-like) structu res. A fter th e extractio n of th e candidate surfaces, they scrolled these surfaces w ith a p la n ar perpendicular on th e m ain axis of th e surface an d th e y com puted th e edge ffow from th e ex trem ity of th e surface tow ards its center. T he divergence of th e edge flow is used to determ ine w hether the candidate surface is generated by a polyp or a fold. T hey applied th is technique on 48 d atasets and their experim ents indicate th a t th eir m ethod achieved 35% specificity a t a sen­ sitiv ity ra te of 100%. T his m etho d was fu rth er advanced by G okturk et al. [99] w hen th ey applied th e random ly oriented triple orthogonal planes a t th e location of each candidate surface. T h ey applied this approach to sam ple th e sphericity of

th e can d id ate surface based on th e fact th a t any random p lan ar slicing through a spherical surface will generate a circle. T he rep o rted experim ental results indicated th a t th ey achieved 69% specificity a t a sensitivity ra te of 100%. No detailed analy­ sis w ith respect to th e size of th e polyps is provided. W ang et al. [40] sta te d th a t th e inclusion of m orphological and te x tu re features can reduce 10 tim es th e false positives w hen com pared to th e stan d ard shape-based approach. W ang et al. [100] com bined th e te x tu re features and global cu rvature for au to m atic polyp detection and shows 100% sensitivity for > 10mm w ith 2.0 false positive p er dataset. Jerebko et al. [41] employed a m ultiple neural netw ork classification scheme to achieve a 36% reduction in F P s and a 20% reduction in false negative (FN )detection. Later, Jerebko et al. [101] employed a su p p o rt vector m achines com m ittee classification scheme to achieve 81% sensitivity w ith 2.6 false positive per d ataset. Iordanescu et al. [102] developed a rectal tu b e detection m etho d th a t was applied to reduce th e F P s generated by th e rectal tu b e. Li et al. [67] proposed m ethod employed different geom etric features such as m axim um polyp radius calculated from th e m inim um curvature, m inim um polyp radius derived from th e m axim um curvature, candidate surface area, roundness of th e can d id ate surface and elongation factor for classifi­ cation of th e can d id ate surfaces into polyps or folds. T heir m ethod achieved 90% sensitivity w ith 2 F P s per d ataset.

All th e above m entioned C A D - C T C techniques show 100% or close to 100% sensitivities in th e detection of polyps > 10mm, while th e sensitivities in th e detec­ tio n of polyps in th e range [5 — 10)m m vary from 70% to 95%. The rep o rted false positive rates vary from 2.0 to 90 p er d ataset. A m ong all th e developed C A D -C T C techniques, Y oshida et al. [37, 92] and Kiss et al. [38] m ethods show best results for sensitivity and false positives incidence per d atase t. Y oshida et al. technique achieved a sensitivity of 100% p er p atien t w ith 2.0 false positive. B ut it is w orth noting th a t th e sensitivity d ropped to 90% w hen it was presented as per polyp. Also th e polyps sm aller th a n 5m m were com pletely ignored in their evaluation. Kiss et al. [38] m ethod shows 90.90% sensitivity for polyps > 9m m and 100% sensitivity for colorectal tum ors w ith a false positive rate of 2.48. T heir m ethod shows 33.33% sen­ sitiv ity for polyps < 6m m and 85.70% sensitivity for polyp between 6 — 9m m where th e d a ta used in th e ir experim ents has been acquired w ith 0.8mm reconstruction interval. It is also useful to n ote th a t b o th Kiss et al. and Y oshida et al. C A D -C T C techniques evaluated th e difference in th e geom etrical shapes between polyps and

folds.

In th is thesis these geom etrically-driven approaches will be fu rth er advanced by developing a num ber of CAD polyp detection techniques w here th e discrim ination betw een polyps and folds is perform ed using th e features th a t sam ple th e morphology of th e local 3D d ata. All th e proposed polyp detection m ethods employed different features derived from th e colon wall in order to classify optim ally th e candidate surfaces into polyps and folds. T he first polyp detection scheme called geom etrical fitting approach evaluates th e discrim inative power of th e features calculated from th e colon surface using least square approxim ation (ellipsoid, sphere, plane) in order to perform polyp identification. T he second m ethod uses th e statistical features derived from th e colonic surface. T h e th ird m eth o d analyses th e 3Decibel (3dB) atte n u a tio n on th e surface variation curve and surface norm al concentration for polyp detection.

3 . 1

G e o m e t r i c a l F i t t i n g A p p r o a c h

F igure 3.1, gives an overview of th e proposed algorithm . In th is section, th e seg­ m en tatio n, polyp surface generation and feature ex tractio n phases of th e algorithm are discussed in detail.

3 D V o lu m e D a ta _{P o ly p S u rfa c e D e te c tio n}S e g m e n ta tio n a n d A u to m a tic C o lo n S e g m e n ta tio n _{o r} M a n u a lly s e e d e d s e g m e n ta tio n G e o m e tric a l F e a tu re E x tra c tio n T ra in D a ta T e s t D a ta N o rm a liz e d N e a r e s t N e ig h b o u rh o o d C la s s ifie r

1

D e c is io n : P o ly p , n o n -p o ly p

Figure 3.1: Overview of th e G eom etrical F ittin g C A D - C T C system.

3.1.1 S egm en tation

C T C images provide high co ntrast between th e gas and colon surface. Using a region growing [69] algorithm th e gaseous region can be segm ented successfully. Som etim es rem aining residual m aterial and w ater can create collapses in th e colon and th e region growing algorithm m ay require m ultiple seed points to segm ent th e

entire colon. The developed CAD-CTC system provides both the manual assisted

In document Tus gorras de natación. facebook/amdetna (página 46-55)