Nguygn Thanh Tmng vd Dtg Tgp chi KHOA HOC & CONG NGHE 162(02): 8 1 - 8 6
SIJ' D T J N G M A N G N O - R O N N H A N T A O T R O N G V I E C K H U " N H I E U A N H C T Nguyin Thanh T r u n g ' , Phgm Thi Huo'ng Trudng Dgi hgc Cong ngh? thdng tin vd Truyin thong - BH Thai Nguyen T O M T A T
Trong bai bao nay chiing toi d^ xudt mflt phuong phdp khii nhilu hi€u qua ap dung cho anh chgp cit lap (CT), Trong phuang phap dugc de xuat, mpt anh dugc phan tach th^nh t6ng ciia cac thanh phin tin so bao gom: thanh phan tan so thap, thanh phan tan so trung binh va thinii phan tan s6 cao. Khir nhilu dugc thyc hiSn thfing qua viSc udc lugng cac thanii phin tin so nay. Th^nh phan tan s6 trung binh dugc udc lugng dua tren bieu dien thua, sau do thanh phin tin s6 cao dugc udc lugng bing viec sii dyng mang na ron nhan tgo. Viec udc lugng tren can sii dung din co sd dO li§u dugc xay dyng ty tap cac anh mlu chudn khdng nhilu, nhttng anh nay tuang ty nhung khong giong het anh can khii nhilu. Thyc nghifm cho thiy phuang phap cua chiing t6i bao ton cac chi tiet nho Uong anh tot hon mpt so phuong phap hien dai gan day cave djnh lugng va dinh tinh, Tu khoa: Khu nhiiu anh, mgng na-ron nhan tao, biiu diin thua, dnh chup cat lap, tia X GIOI T H I E U
Ky thuat hinh dnh C T (Computed tomography) tgo ra anh chyp cat Idp bang viec quet chiim tia X qua c a the ngudi theo lat c i t ngang. A n h chyp cat Idp C T cd vai tro rat quan trpng ddi v d i chan doan mot sd benh.
Trong qud trinh tao anh C T , dac biet la ky thugt sii dung liSu lugng tia X thap, su x u i t hien ciia n h i l u la mgt trong cac y6u to lam c h i t luong hinh dnh bj giam [1], [2] va cd the d i n tdi viec gidra sy chinh xdc trong viec chan dodn benh.
Nhieu trong anh C T lien quan tdi s6 lugng photon tia X gdp p h i n tgo nen diem anh. Do vgy, ta cd t h i giam nhieu bang viec tang lieu luong tia X khi quet. Khi tang lilu lugng tia X dflng nghTa vdi v i e c tang miic dg gay hai tdi siic khoe ciia ngudi benh, M g t giai phap khac d u g c uu tien la phdt trien cdc p h u a n g phap khii n h i l u dii mgnh d l cd dugc hinh anh c h i t lugng dam bao vdi lieu lugng tia X thap n h i t cd the. Vi vgy, m a c dii khii nhieu dnh la bai todn kinh d i l n nhung dfli vdi anh y t l noi chung, anh C T noi rieng van cdn nhieu thach thiic c i n d u g c giai q u y l t .
D I khii n h i l u dnh cd r i t n h i l u each t i l p cgn khac nhau mfli each t i l p can co iru, nhugc diem rieng va cd phgm vi ap dyng khac nhau
' Tel. 0987 843338, Email: [email protected] v,
[3]. Cdch don gian nhat de khti' nhieu la su dung cac phep Igc cfl dien trong mien khdng gian n h u cdc phep lpc Gaussian, Wiener,.,.
[4] Nhiing bp Igc nay cho ra k i t qua rat nhanh va tdt vdi nhfl'ng vung dong nhat nhung Igi lam md cdc cau true bien va cac chi tiet nho.
Trong c h i n dodn benh su' diing dnh CT, nhung chi t i l t nhd co the chua dung nhung thong tin quan trgng ve benh tat. Do vgy, tim ra mdt phuong phap khu nhieu sao cho nhieu dugc loai bd ma van bao ton duoc nhieu cac chi t i l t nho la mpt trong nhung thach thii'c khong nho va co y nghTa.
Cho tdi hien tai, co mot s6 p h u a n g phap khii' n h i l u chii trpng d i n viec bao tfln nhiing chi t i l t nhd [5], [6], [7], [8], Ben canh cac han che ve do phiic tap tinh toan va yeu cau chat che trong xay dyng c o sd du- lieu, nhii'ng p h u a n g phap nay cho ket qua day trien vpng.
Chiing tdi de xuat mpt phuong phap khii' nhilu trong do anh dugc khu nhieu dya tren vi?c khu nhilu d6i vdi timg thanh phan tan s6, ket hgp viec sii' dung bieu dien thua va mgng na-ron nhan tgo. Trong dd, chung toi tach anh d i u vao thanh cdc thanh p h i n tin sd t h i p , trung binh, cao, roi tien hanh udc lugng tii'ng thanh phan nay, Thanh phan tan so thap dugc l i y tir anh nhilu sii dung bp lpc thflng thap Gaussian, thanh phan tan so trung binh dugc udc lugng bang each su dung bieu dien
Nguyin Thanh Trung vd Dig Tgp chi KHOA HOC & CONG NGHE 162(02): 81-86
thua, thanh phan tan s6 cao udc lugng thdng
qua mang na-ron nhan tgo, viec udc lugng hai thanh phin tin so trung binh va cao dugc thyc hi^n dudi sy trg giiip cua co sd dii lieu gom cac cap mieng anh tan sd trung binh, tin sd cao: (?„,, Ph). Co sd dii lieu nay dugc tgo ra tii tap cac hinh anh mau chat lugng tot va dugc xem nhu khong nhieu,
PHUONG PHAP DE XUAT M& hinh cua anh nhieu
Nhieu trong anh CT dugc chiing minh la cd the xip xi bing phan bo Gaussian [I]. Do vgy, neu gpi Y la anh bj nhieu thi ta cd:
Y = x + n (I)
trong do X Id anh khflng nhieu (can udc
lugng) T| la bien ngau nhien tuan theo phan bd Gaussian co trung binh bang 0 va phuang sai
a^. Miic do nhieu nang hay nhe tuy thupc vao gia tri cua phuang sai.
Phan tach anh
Viec phan tach cdc thanh phan tan sfl [8] dugc thyc hien thong qua cdc bp lpc trong miln khdng gian cu the nhu sau:
Y' = LPF(Y) (2) Y™ = BPF(Y) (3) Y"" = Y - Y" - Y' (4) Trong do LPF(Low Pass Fiher), BPF (Baad Pass Filter) tuang ling la bp lpc thdng thap va bp lpc thdng dai. Cdc chi sd I, m, h tuang ung vdi low (thip), middle (trung binh) va high (cao). Trong thyc nghiem, chung toi sii dung cac bg lpc thdng thap Gaussian ket hop vai toan tii trii de thyc hien viec phan tach mot anh ra thanh cac thanh phan tan so Y', V", Y\
Hlnh \. Phdn tdch dnh thanh cdc thdnhphdn tan so thap, trung binh, cao Y tudng phirong phap
Vi^c khu nhieu anh thyc chat la viec udc lugng X tCr Y trong md hinh (1). Vdi each phan tach anh thanh cac thanh phan tan sd thap, trung binh vd cao, ta co the khir nhilu thdng qua vi?c irac lugng tung thanh phin tin so ciia anh diu ra. Vdi gia thuylt nhieu chii yeu tap trung d Y"", Y*" va it tap trung d Y' nen chiing tdi tap trung vao viec khic nhilu Y"" vd Y*", thyc chit la thyc hien ii6c lugng X™ va X tir Y dudi su ho trg cua co sd dii lieu dugc tgo ra tir nhiing anh chat lugng tk dugc xem nhu khdng nhieu. Vdi each tilp can nay, ket qua ciia anh diu ra sau khi khu nhilu la;
X=Y^ + X^+X'* (5).
D I udc lugng X"", X** cd nhilu each, trong phuang phap MRFD [8] , X"" = V" va X'' dugc uac lupng thdng qua trudng Markov ngau nhien, md hinh nay khd phiic tap. Trong phuang phap cua chung tdi, X" dugc udc lugng dya tren bieu diln thua roi sir dung kit qua nay dl udc lugng X'' thdng qua mang na ron nhan tgo. Phuang phap gdm cd cac budc sau:
Birdc 1: Xay dyng ca sd dii li^u
Bu^c 2: Udc lugng thanh phan tin sd trung binh Bu-dc 3: Udc lugng thanh phin tin so cao.
Budc 4: Tdng hgp anh diu ra Noi dung phirong phap Budc 1: Xay dung ca sa AH lieu
Vdi mdi logi anh nhieu diu vao, chiing tdi chpn cac anh hrong tu nhu anh nhilu (chup d nhiing v[
tri gan vdi vi tri anh nhieu). Nhung dnh nay dugc xem la anh chuin khdng nhilu. Sau dd, chiing
82
Nguyin Thanh Trung va Dtg Tgp cht KHOA HQC & CONG NGHE 162(02): 81- tfli tach cac anh tr6n thanh cac thanh phin tin
sd: thap, trung binh va cao theo each o muc Phan tich anh. Thanh phin tan s6 cao va thanh phan tan s6 trung binh dugc chia thanh cac mieng anh vudng nhd goi nhau, cung kich thudc rdi sip thanh cgp (v-^.n^ dl hinh thanh nen co sd du lieu (P^, Ph) bao gdm cac cap mieng tin s6 trung binh-tin s6 cao. Mdi milng sau dd dugc luu trii dudi dgng vec to N chieu vdi:
N = (so ddng diem anh ciia mieng)*(sfl cot diem anh ciia mieng).
Co sd du lieu nay se dugc su dyng trong viec udc lugng X™ va X"" cua anh dau ra tir Y™ ciia anh nhieu.
Budc 2: U'ffc lirgug X"
Viec tim X™ dugc thuc hiSn theo tirng milng nhd, nghia la tim timg mieng i f tuang ling vdi mieng y™ trong Y" , cy the nhu sau:
a) Anh nhieu dugc phan tach ra thanh cac thanh phan tan sd thap, trung binh, cao.
Thanh phan tan so trung binh dugc chia thanh cac tap M mieng gdi nhauj'f ( i - 1,2,...M).
Cac mieng cd kich thudc bang vdi kich thudc cua mieng trong ca sd dii lieu.
b) Vdi mfli mieng yT > ta tim ca sd dfl li?u con Df (K lan cgn gan nhat cua y f ) trong Pm, sau dd ^ dugc tinh thdng qua bieu dien thua, cu the nhu sau:
i f = flf • a (6) trong dd:
a = arg min
l\\Dr«-
(7) sao cho |fflo < i vdi L < if.
c) Ghep cac milng 2 f (i = 1, 2,...M) vdi nhau ta se thu dugc JT".
Budc 3: Vffc lugng X""
Vi?c uoc lugng thanh phan tan so cao X cQng dugc thyc hien theo timg mieng nhd.
Chung tdi sii dyng mgng na-ron da lap truyen thing MLP [9], [13] de udc lugng tirng milng nhd x f tu S f da tim dugc d budc 2c. Quy trinh nhu sau:
a) Huin luyen mgng MLP
b) Sir dung mang da huan luyen de tim ra x^
bang cdch cho 2 f qua mang viia huan luyen.
Cau tnic chung ciia mot mang MLP nhu hinh 2.
Hinh 2. Cdu triic chung ciia mgng MLP
Co sd dii lieu diing trong viec huan luyen mang la (Pm, Ph) dugc xay dyng d budc 1.
Viec huan luyen co the dugc thyc hien doc Igp tir trudc, trong tru'dng hgp nay thdi gian danh cho viec udc lugng thanh phan tan sfl cao cua anh dau ra se giam di dang ke.
Sau khi udc lugng dugc tap cac mieng x^
tuang li'ng vdi cac mieng 5^f ( i =1, 2, ., M), thanh sd cao X^ ciia dnh dau ra dugc tim ra bang cdch ghep cdc mieng nay vdi nhau, Budc 4: Tong hgp anh dau ra
Anh dau ra dugc tflng hgp tir cac thanh phan tan so thap, tan so cao, tan sfl trung binh theo phuong trinh (5), trong do X^ = Y^
MO PHONG VA DANH GlA KET QUA De danh gia hieu qua cua phuang phap chiing toi tiin hanh mpt sfl thuc nghiem tren anh CT vimg phdi (hinh 3). Trong thuc nghiem nay, dnh nhieu dugc tao ra bang each cpng bien ngau nhien Gaussian vao anh gflc (xem nhu khdng nhieu). Cac miic nhieu duoc lya chgn vdi gia tri a lan lugt la 10, 20, 30. Nhieu logi kich thudc mieng khac nhau dugc thiet lap khi lam thuc nghiem: 9x9, 11x11, 13x13, 15x15, 17x17.
Mgng MLP (hinh 2) dugc lya chpn vdi cau tnic 2 Idp. Sd lugng dau vao, dau ra ciia mgng dugc lya chpn theo kich thudc mieng anh, Nhieu bp tham sfl khac nhau (so no ron an, loai ham kich hogt) dugc thiet lap cho viec huan luyen mang.
83
Nguyin Thanh Trang vd Dtg Tap chi KHOA HOC & CONG NGHE 162(02): 81 -
Hinh 3. Anh CT viing phot khong nhieu Tuong ling vdi logi anh dau vao tren (CT
vimg phoi) mdt so anh tuang tu nhimg khong giong hpt vdi anh tren dugc lya chgn de xay dyng CO sd dii lieu
Hinh 4. Mdt so dnh dimg xdy dung CSDL
De thuc thi (7) chiing tdi dya tren thuat toan OMP [ 10]. Chi so SSIM(Structural SIMilarity) [12] duac lya chpn dl danh gia dinh lugng. Chi so nay nim trong doan [0,1]
gia tri SSIM cang Ion thi sy tuang dong ve cau tnic giiia hai anh cang Idn. Phuang phap ciia chiing tdi duoc so sanh vdi cac phuang phap phuong phap NLM [3], Wiener [4], TGV [II]. De tinh toan SSIM chiing tfli sii dyng chuong trinh tinh SSIM tai ve tir trang web cd dia chi
htlps.//ece uwaterloo.ca/~z70wang/research/ssim/.
Bang dudi day la mdt phan trong cac kit qua thuc nghiem chiing tfli dS lam.
CT
Ph&
a 10 20 30
WN 0 8758 0 7493 06364
NLM 0.9300 0 8458 07590
TGV 09035 0 8681 06145
ppaT
Kuat 09227 0.8835 0*MS
Quan sat ket qua trong bang dl dang nhiiii thay hau het chi so SSIM doi vdi phuong phap ciia chung tfli cao ban SSM cua cac phuang phap khac, dac biet doi vdi miic nhieu cang cao phuang phap ciia chiing toi cang td ra hipu qua. So lieu nay cung phan nao chiing td phuang phap cua chiing toi bdo ton cau tnic anh tot hon.
De danh gia ve mat dinh tinh chiing toi duara
trong bai bao nay ket qua khu nhieu anh CT
viing phoi trong truong hgp miic nhieu tuong
ling vdi CT = 20.Nguyin Thinh Tmng vc Dig Tgp chl KHOA HQC & CONG NGHE
Hinli 5f. Plniun^phap cle xiidl Quan sat cac anh dau ra tii phuang phdp ciia chiing tdi va cac phuang phdp khac ro rang ket qud cho bdi phuong phap dugc chiing tfli de xuat bdo toan dugc cdc chi tiet nhd t6t hon so vdi cac phuang phap cdn lai.
KET LUAN
Trong bai bao nay, chiing toi thuc hi^n khir nhieu anh CT bang vi?c ket hgp bieu dien thua va mgng na ron nhan tgo de udc lugng thanh phin tin sd trung binh vd thanh phan tan sd cao ciia anh, nhung thanh phan nay chiia nhieu thdng tin ve cau true va cdc chi tilt nhd ciia anh. Ket qud thyc nghi?m cho thay phuong phdp ciia chiing tdi kbit nhieu hi^u qua trong khi bao ton dugc cdc chi tiet nho tot hon mgt so phuang phdp. Trong tuang lai chiing toi tien hanh thCr nghiem danh gia vdi ca sd dii lieu phong phii ban va tren cdc ciu true mgng no-ron nhan tgo khac dong thdi khao sat sy anh huong ciia tat cd cdc tham so din ket qua ciia phuong phap.
Bai bao san pham ciia de tai cd mS so T2016-07-23 dugc tai trg bdi kinh phi cua tru-dng Dgi hgc Cong nghe Thdng tin va Truyen thdng - DHTN. Nhdm tac gia chan thanh cdm on sy tai trg cua quy Trudng.
TAI LIEU THAM KHAO 1, H. Lu, l.-T, Hsiao, X, Li. and Z Liang, "Noise properties of low-dose CT projections and noise treatment by scale transformations," in Nuclear Science Symposium Conference Record, vol, 3, 2001, pp, 1662-
1666.
2. L. Yu, X. Liu, S Leng, J. M. Kofler, J. C, Ramirez- Giraldo, M Qu, J. Christner, J, G, Fletcher, and C, H.
McCoUough, (2009) "Radiation dose reducUoti in computed tomography: techniques and fiiture peispeciive,"
Imaging in Medicine, vol I, no, 1, pp. 65-84.
3 A, Buades, B, Coll, and J.-M. Morel, "A review of image denoising algorithms, witli a new one," SIAM Journal on Multiscale Modeling and Simulation, vol 4, no. 2, pp. 490-530, 2005
4. J. S. Lim, (1990) Two-Dimensional Signal and Image Processing. Upper Saddle River, NJ, USA' Prentice-Hall, Inc.
5 D, H, Trinh, M Luong, J.-M, Roccliisani, C. D.
Pham, H. D Pham, and F. Dibos, (2012) "An optimal weight method for CT image denoising," Journal of Electronic Science and Technology, vol, 10, no, 2, pp.
124-129.
6. D. H. Tnnh, M, Luong, J.-M, Rocchisani, C. D.
Pham, and F. Dibos, (2011) "Medical image denoising using kernel ridge regression," m I8"'!EEE Int Conf on Image Processing (ICIP). IEEE, pp 1597-1600, 7. D. H. Trinh, M Luong, F. Dibos, J -M. Rocchisani, C, D, Pham, and T, Q, Nguyen, (2014) "Novel example- based method for super-resolution and denoising of medical images," IEEE Transactions on Image Processing, vol. 23, no 4, pp. 1882-1895.
8. D, H, Tnnh, N Linh-Trung, and T-T Nguyen, (2014) "An effective examplebased denoismg method for ct images using maikov random field," in IEEE Inl.
Conf on Advanced Technologies for Communications.
IEEE. pp. 355-359.
9. H.B. Demuth and M. Bealc, Neural Network Design, PWS Publishing company, 1996 10 Y. C. Pali, R. Rezaiifar, Y. C P. R. Rezaiifar, and P 5. Krishnaprasad, "Orthogonal matching pursuit:
Recursive function approximation with applicalions to wavelet decomposition," in Proceedings of the 27 th Annual Asilomar Conference on Signals. Systems, and Computers, 1993, pp, 40--14
II, K Bredies, K Kunisch, and T Pock, (2010) "Total generalized variation," SIAM J. on Imaging Sciences, vol. 3, no. 3, pp. 492-526,,
12 Z, Wang, A C Bovik, H R Sheikh, and E, P, Simoncelli, (2004) "Image quality assessment' from error visibility to structural similarity," IEEE Transactions on Image Processing, vol. 13, no 4, pp.
600-612, Apr.
13. hllps,//www,mathworkscom/matlabcgnlral/filcexch ange/2654-
netlab?reque5ledDomain-www,mathwQrkscom
Nguyin ITianh Trung va Dig Tgp chi KHOA HQC & CONG NGHE 162(02): 81 -
S U M M A R Y
A P P L Y A R T I F I C I A L N E U R A L N E T W O R K TO D E N O I S E C T I M A G E Nguyen Thanh Trung*, Pham Thi Huong College of information and Communication Technology - TNU In this paper, we propose an efficient denoising method for CT image. In our work, an image is decomposed to three frequency bands, namely low- band, mid- band and high-band.
Thus denoising can be performed by getting these frequency bands. The mid-band is estimated based on sparse representation, and then the high-band is estimated by applying artificial neural network. The estimating uses database constructed from standard sample images that similar but not identical the input image. The experiment results show that our method is better than some other state-of-the-art methods both in subjective and objective comparison.
Keyworks: denolse, sparse representation, artificial neural network. Computed Tomography, Xray
Tel: 0987843338. Email: [email protected] vn
