Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

• Diagnostic

Radiology
Automated Diagnosis of Congenital Dislocation of the
Hi p1
Masahlro Endo, M.Sc., Takeshl A. IInuma, Ph.D., Yolchlro Umegakl, M.D.,
Yuklo Tateno, M.D., Hltoshl Tanaka, B.Sc., and Kellchl Tsuchlya, M.D.

A study of the automated diagnosis of luxatio coxae congenita (LCC) was performed to es-
tablish an automated screening system for hip radiographs. A specific parameter with high
diagnostic accuracy was selected by the use of information theory. Pattern recognition
programs were written to extract the parameter from digitized hip radiographs. Parameter
measurement was performed on 10 radiographs and compared with manual measurements.
Results showed good agreement with measurements obtained by specialists. This study
confirms the possibilities of automated diagnosis of LCC.

INDEX TEAMS: Bones, abnormalities • Computers • Hip, congenital dislocation,


4 [4] .146 • Hip, dislocation
Radiology 122:25-30, January 1977
• •
HE INCIDENCE of luxatio coxae congenita (LeC) among
TJapanese infants is high compared to other nations.
Early diagnosis is required to achieve an effective cure.
Radiographic survey is the most reliable method and is
widely used in Japan. However, this method presents
, several difficulties: (a) the gonads are exposed to radiation,
and (b) it requires an inordinant amount of time. An auto-
mated, low-dose, diagnostic system is urgently needed.
We are developing a new low-dose, automated, x-ray
imaging system able to diagnose LCC. In this paper, a
preliminary method of automated diagnosis is described
as a part of the whole system. A specific parameter with
high diagnostic accuracy was selected by the use of in-
formation theory. Pattern recognition programs were Fig. 1. A typical radiograph of an infant's hip joints. The center
written to extract the parameter from digitized hip radio- shadow is due to shields.
graphs. Parameter measurement was performed on 10
radiographs of 3-month-old infants and were compared Table I: The Result of Orthopedists' Readings
with manual measurements. The results showed good True False Information
agreement with measurements obtained by specialists. Doctor Positive Positive Content*
# (%) (%) (bits)
QUANTITATIVE EVALUATION OF LCC DIAGNOSIS LCC SPECIALISTS
1 73.1 0.0 0.344
It is important for the computerized diagnosis to es- 2 65.4 0,6 0.271
3 88.5 6.3 0.294
tablish quantitative standards (rates of correct diagnosis). 4 84.6 0.0 0.419
To do this, one must first ascertain the diagnostic reliability 5 84.6 4.6 0.295
0.325 (Average)
of the specialists.
GENERAL ORTHOPEDISTS
Diagnostic reliability of LCC specialists: In order to 65.4 10.3 0.130
6
establish the diagnostic reliability of LCC specialists, we 7 76.9 1. 7 0.302
performed a study in which orthopedists read 100, con- 8 65.4 2.9 0.215
9 69.1 1.1 0.274
firmed hip radiographs. These were randomly selected 10 80.8 1. 7 0.328
from radiographs (Fig. 1) taken during mass examinations 11 76.9 5.2 0.240
0.249 (Average)
by the department of orthopedics and include 174 normal
hip joints and 26 abnormal joints. Final diagnosiswas made *The information content of the input side = 0.557 bits.
by further clinical follow-up. These radiographs were read
by five specialists and six general orthopedists. Readers centages. We evaluated these percentages by introducing
classified them into normal and abnormal groups. From an element of information theory: The larger the value of
'these we calculated true positive and false positive per- the reader's information content; the smaller the error in

1 From the National Institute of Radiological Sciences (M. E., T. A. I., Y. U.), Anagawa 4-9-1, Chiba-shi, Japan; and the Departments of Radiology
(Y. T., H. T.), Chiba University Hospital, and Orthopedics (K. T.), Chiba University School of Medicine, Inohana 1-8-1, Chiba-shi, Japan. Accepted
for publlcatlon in May 1976. elk

25
26 MASAHIRO ENDO AND OTHERS January 1977

X=c/b
Z=X+Y
y:=a./y
A a= (0 I D/OO ') XIOO B
Fig. 2. A. The diagnostic parameters of LCC: h, d, and a are called Hilgenreiner's parameters. A ratio, a = (0'0/00')
X 100, is called Iino-Imada's parameter. The length of HT is called Miki's parameter. Angle (3 is called Uziie's parameter.
B. Ratios X, Y, and their sum Z, are called Nozaki-Nagai's parameters. All parameters are measured for right and left
hip joints.

his interpretation (see Appendix for further detail). The


30
average value of the information content of the five spe-
cialists is 0.325 bits and is larger than that of the general
orthopedists (TABLE I). It is preferred that the value of the
o
r.:l
Cl
information content of the quantitative diagnosis not be
N
...... 20 smaller than 0.325, if possible.
til
E-<
Z Choice of a parameter for LCC diagnosis: Several di-
~
o
o agnostic parameters have been obtained by geometric
measurement of hip radiographs (1). Nine parameters are
10
ABNORMAL presented in Figure 2, A and B. They can be measured by

,,....,~/
the feature points on the hipbone borders. Values of the
, , , nine parameters were measured by manual bone-edge
M \ ,~
tracing on the same 100 radiographs read by the ortho-
20 3D 40 50 60
pedists. Figure 3 shows two histograms, one for normal
UZIIE'S PARAMETER (DEG.)
cases and another for abnormal cases, obtained by the
Fig. 3. The histograms of Uziie's parameter. The solid line corre- measurement of Uziie's angle «(3).
sponds to the normal group and the broken line to the abnormal

" .
group. Diagnostic levels of the parameters were compared by
the Receiver Operating Characteristics (ROC) curve
method (2). The ROC curve is, essentially, a graph of the
100 e " .

(J;
conditional probability of a true positive response vs. the
a"",'~.: ~ ······ conditional probability of a false positive response in a
,,' .",/"" a detection experiment.
, ••'/
I If the distribution of a parameter is given for both normal
I' . ..: and abnormal cases, an ROC curve can be drawn for that
//' parameter by shifting an operating point which separates
,, ::
, !
normal and abnormal groups. Since for each of nine LCC
,:,!
,;
t: parameters, histograms of distribution for normal and
If
:i abnormal cases have been constructed from manual
::
~
measurements of 200 hip joints, the ROC curves of these
t
t
(
parameters can be obtained. It can be seen in Figure 4 that
t Uziie's angle «(3) gives the curve which is closest to the
point where true positive probability = 100 % and false
positive probability = 0 % , and therefore has the highest
diagnostic reliability. Six other parameters not shown in
I
Figure 4 are all inferior to Uziie's parameter in terms of the
o 50 100 ROC curve.
FALSE POSITIVE (%)
The ROC curve is very useful for the evaluation of di-
Fig. 4. ROC curves of the diagnostic parameters of LCC.
agnostic reliability but is not strictly quantitative. For ex-
The dotted line shows an ROC curve obtained by the diagnostic
parameter a. The broken line and solid line show ROC curves ample, if two ROC curves cross each other, two situations
corresponding to parameters a and (3, respectively. can not be compared by this method alone. Metz (3) has
Diagnostic
Vol. 122 AUTOMATED DIAGNOSIS OF CONGENITAL DISLOCATION OF THE HIP 27 Radiology

100 r-------~_-------____.
solved this problem by relating the information content to
the ROC curve.
Any point on a ROC curve indicates a pair of values of
true positive probability and false positive probability. From
the relation mentioned in the Appendix, one can calculate
the information content for each point on the ROC curve ~
and regard their maxima as the measure of diagnostic
~
accuracy of the curve. ~ 75
H
TABLE II shows the maximum information contents of til
0
Il<
nine LCC parameters. Uziie's parameter gives the largest
~
value of all and, therefore, shows the highest level. ~ )( X
E-!
Comparison of the quantitative diagnosis with special-
ists: False positive probabilities and true positive
probabilities shown in TABLE I are plotted in Figure 5 which
also shows the ROC curve obtained by Uziie's angle ({3). 50 ~ ____IL..- ___I

Since the ROC curve passes among the circles which o 25 50


show the results of LCC specialists' diagnoses, it is con- FALSE POSITIVE (i)
cluded that the diagnostic reliability of Uziie's angle ({3) is Fig. 5. An ROC curve of Uziie's parameter. Circles
as high as that of the specialists. indicate the results of LCC specialists' diagnosis. Crosses
The same conclusion results from the analysis of the (x) indicate results of general orthopedists' diagnosis.
information contents. From TABLE I we get 0.325 bits for
the average information content of specialists. On the other
~
hand, from TABLE II we get 0.360 bits for the maximum
information content of Uziie's angle ({3, 37.25°). So the
diagnostic level of {3 is as high as that of average LCC
DENSITY
F-TYPE Q
specialists.

COMPUTER PROGRAM FOR AUTOMATED DIAGNOSIS ~

~
OF LCC

Since Uziie's parameter can be measured from the


borders of hipbones, computerized diagnosis of LCC is REFLECTING POWER
made mainly by the automated extraction of those borders. 2 PIX/MM
330X200 8 BIT
This is difficult to do from complex multilevel images.
A general flow-chart of the system is shown in Figure
6. A hip radiograph (Fig. 1) is inverted to the positive
[REj:]
photograph of the same size. It is scanned by the raster MT-REGISTRATION
method and digitized by the photo-electric image input
apparatus using a rotating drum (4), and a reflection Fig. 6. The general flow-chart of the system.
coefficient of each picture cell (pixel) is punched onto the
paper tape. The resolution is 2 pixels/mm, the number of
sampled pixels is 200 X 330 for every film, and the gray Table II: Maximum Information Contents (bits) of the
scale accuracy is 8 bits. The sampled area is 10 X 16.5 Quantitative Diagnosis
cm and contains the main parts of the hip radiograph. h 0.127 {3 0.360
Image data on the paper tape are converted into mag- d 0.130 X 0.315
a 0.159 y 0.266
netic tape by the TOSBAC-3400 at our institute. Reflection a 0.196 Z 0.322
coefficient is also corrected to the film density. These data HT 0.154
are processed by the pattern recognition software, Auto-
mated Diagnosis System of LCe Roentgenograms
(ADSLR-1) (Fig. 7) which consists of the following sub- rameter and is so large that it can not be processed all at
systems: once by our computer. In this subsystem, necessary re-
(a) Limitation of objective regions gions are extracted by the use of the total information of
(b) Extraction of contour line of bone edge the image.
(c) Simplification of contour line Figure 8 illustrates the algorithm by which the objective
(d) Curve tracing of hipbone borders regions are extracted. The spatial signatures proposed by
(e) Extraction of feature points and measurement of Kruger et al. (5) are used in this subsystem. The horizontal
parameters and vertical signatures are computed by summing the
Limitation of objective regions: The input image con- picture elements in the columns and rows of the image
tains elements unnecessary for extracting Uziie's pa- array. The large rectangle in Figure 8 describes a total input
28 MASAHIRO ENDO AND OTHERS January 1977

The vertical signature of left and right areas between


the center dotted line and side lines has a maximum cor-
responding to the Y-cartilage and a minimum corre-
sponding to a line which connects bones. Four reference
LIMITATION OF OBJECTIVE points are determined by the combination of these eight
REGIONS coordinates which gives this maximum and minimum.
Objective regions are given by these four points. Refer-
ence points are located by cross marks (x) and objective
EXTRACTION OF CONTOUR LINE regions as small squares in Figure 8. Every region contains,
OF BONE EDGE
either the upper part of each femur, or the lower part of
each ilium. These are necessary for extracting feature
points for the measurement of Uziie's parameter which
SIMPLIFICATION OF CONTOUR
LINES consists of 64 X 64 picture elements.
Extraction of contour line of bone edge: Many methods
are proposed to extract contour lines from radiographs.
CURVE TRACING OF HIP-BONE Chow (6) proposed the dynamic threshold method to ex-
BORDERS' tract the left ventricle boundary from cineangiograms.
Ballard (7) used dynamic programming to detect tumor
boundaries from a chest radiograph. These methods are
not suitable for our cases.
EXTRACTION OF FEATURE POINTS
AND MEASUREMENTS OF PARAMETER Shikano et al. (8) proposed the wave-propagation
method to extract borders of ribs from chest photofluo-
rograms. As this method seemed very promising, we de-
cided to use it. The details of the method are not mentioned
here, but our figures show how the contour lines are ex-
Fig. 7. The block flow-chart of ADSLR-1. tracted.
Figure 9, A is an original, 8-level printout of the region
described by subsystem (a) and includes the upper part of
the right femur. The longitudinal size is about twice the

!~~~
£
8 j j PHASE-l
transversal size.
When the original printout is smoothed and edge-en-
Vertical Sig,
1:
,g' hanced by spatial differentiation (Fig. 9, B), it shows high
Bnght Dark ill
differential coefficients along the bone edges. In Figure
9, B zero is substituted in the pixel which has a differential
coefficient less than a certain threshold. Figure 9, C shows
the result. In this procedure, one can estimate the domains
which include line segments. The wave-propagation
method is a series of procedures which extracts line
segments from the roughly-described domains. Figure 9,
D shows the result.
Simplification of contour lines: In Figure 9, 0 the
bone-edge borders are roughly extracted, but lines are not
completely thin and there are small noises. The result of
the simplification process (9) is shown in Figure 9, E.
Curve tracing of hipbone borders: Figure 9, E clearly
Fig. 8. Illustration of the algorithm of subsystem (a). shows extracted hipbone borders but contains noise
caused by a muscle shadow. With this subsystem hipbone
borders are visible despite noise lines caused by the an-
region. It is divided into upper and lower halves (- • - .) at atomical structures, and if necessary, can be interpolated
the maximum point of the vertical signature computed from and recognized as objective lines. The system registers
total area. numerical codes which indicate the directions of hipbone
The horizontal signature calculated from the lower half borders (Fig. 10). For example, 1412 is registered in Figure
has three valleys. The valleys at both sides correspond to 10 which deals with the upper part of the right femur. An
the femurs. The sum of film densities for bone becomes actual border is traced with reference to this code. Noise
a valley in the signature because bone has thin density on lines are deleted when they fail to match programed nu-
the radiograph. The center valley is made by the radiation merical patterns.
shields. The horizontal signature of the upper half has Extraction of feature points and measurement of pa-
minimums corresponding to the ilia. rameters: Feature points necessary to calculate Uziie's
Diagnostic
Vol. 122 AUTOMATED DIAGNOSIS OF CONGENITAL DISLOCATION OF THE HIP 29 Radiology

REGISTERED PATTERN
TRACE OF ACTUAL PATTERN

INTERPORATE IF THERE IS
A CUTTING OF LINE
DIRECTION CODES

Fig. 10. Illustration of the algorithm of subsystem (d).

RESULTS OF ADSLR-l /

,...<.--- Manual
// - Computer
// • Feature Points
.----+-------,

Fig. 11. The result of ADSLR-1. The dotted lines are manually-
traced borders of hip bones (including shields). The solid lines are those
extracted by ADSLR-1. Four squares show the objective regions de-
scribed by subsystem (a).

Table III: The Results of ADSLR-1 (degrees)

Case No. ADSLR-1 Manual 0

L 33.1 32.5 0.6


556 R 40.6 40.5 0.1
L 23.3 22.5 0.8
K199 R 21.1 24.0 2.9
L 31.2 31.5 0.3
Fig. 9. A-E. Example of 731 R 36.7 36.5 0.2
conversion of a radiograph L 23.8 25.0 1.2
514 R 29.3 28.0 1.3
into line figure printouts. See
text for details. L 33.7 35.0 1.3
563 R 29.1 29.5 0.4
L 27.8 29.0 1.2
K246 R 29.9 31.5 1.6
L 25.4 26.5 1.1
518 R 23.0 25.0 2.0
L 38.5 39.5 1.0
445 R 42.1 44.0 1.9
L 36.7 38.0 1.3
632 R 35.1 34.0 1.1
L 31.7 31.0 0.7
parameter are easily detected from the extracted borders 511 R 28.4 29.0 0.6
of hipbones. Figure 11 shows extracted borders of bone- Land R mean left and right side of hipbones, respective-
edge and detected feature points as well as manually ly. 0 is difference between parameter values calculated by
traced borders. Uziie's parameter is calculated from the ADSLR-1 and those measured by man. Cases 556 R, 445 L,
and 445 R are abnormal.
coordinates of these points.
Results of ADSLR-1: Table III shows calculated values
of the parameter vs. those measured manually. They are CONCLUSION
in close agreement; the mean difference being about 1.00 .
This value is almost within the range of measurement error In this work, we have established the quantitative
and is negligible for the difference between average values method for diagnosing Lee from hip radiographs. It uses
of two groups (Fig. 3). a parameter (Uziie's angle) that can be measured by simple
30 MASAHIRO ENDO AND OTHERS January 1977

Table IV: The Decision Matrix of Doctor #1 image technology groups of the Inst Electronics Comm Eng Japan, IT
72-24, Oct 1972
Dr. 1's diagnosis 10. Tateno Y, Tanaka T: Low-Dosage X-Ray Imaging System
N A Total Employing Flying Spot X-Ray Microbeam (Dynamic Scanner). Accepted
confirmed diagnosis n 174 o 174 for publication in Radiology.
a 7 19 26 11. Mochizuki S: On Mathematical models of x-ray diagnosis.
Total 181 19 200 (Japanese) Nippon Acta Radiologica 22:24-36. Apr 1962
N, A, n, a are defined in the text, 'page 30.

APPENDIX
anatomical points. Since the criterion in choosing the
Information content was first introduced from information theory by
parameter is its diagnostic accuracy, it is desirable that
Mochizuki (11) as the measure of the efficiency of the radiographic
the accuracy of the parameter be as high as that of the diagnosis system. Metz (3) also proposed it independently in relation
specialists. By the analysis of ROC-curves, we have shown to ROC curve analysis.
that Uziie's angle satisfies this requirement. Moreover, we The information content of the message can be written as
have developed the software system which extracts Uziie's
angle and have thus confirmed the possibilities of auto-
I = 1;~ p(Yj1Xi) p( Xi) log, [p(Yj 1Xi) /p(Yj) ] (A-1)
mated diagnosis of LCC. If a highly efficient image-input ij
apparatus and a sufficiently large computer were com-
bined with more software improvements, it would be
practical to do automated screening of hip radiographs on expressed in bits, where XI and Yjare state variables of input and output
sides, respectively; p(Xj) and p{Yj) are associated probabilities of these
a large scale.
variables, and P(Yi IXi) is the conditional probability of output Yj given
Dose reduction: We are developing a new method
input Xi. This value can be interpreted as the average reduction of un-
designed to reduce radiation exposure to the gonads (10). certainty after receiving the output message.
An x-ray pencil-beam is used to track around the contour The information content of the message cannot exceed the infor-
of the hipbones. This computer-controlled beam is gen- mation content of input side H(x), where
erated by collimating the rays through a pinhole. With the
use of edge-tracking software, a completely automated
diagnosis requiring minimal exposure will be realized. This H(x) = -1; P(Xi) log2P(xj) (A-2)
i
new method will use the same parameter presented in the
present paper.
expressed in bits.
ACKNOWLEDGMENTS: We thank Dr. Mikio Takagi, associate pro- To apply this idea to radiology. we must regard a series of confirmed
fessor, Tokyo University, for his aid in radiograph digitization; and Mr. radiographs as the input message. In this case. input states are a
Kenziro Fukuhisa of the data processing laboratory for his computer confirmed normal case and a confirmed abnormal case. If we use n for
programming assistance. the normal case and a for the abnormal case, their probabilities can
Division of Clinical Research be written p(n) and p(a)..The radiologist's diagnosis is regarded as the
National Institute of Radiological Sciences output message. Output states are a diagnosed normal case and a
Anagawa 4-9-1 diagnosed abnormal case with probabilities p(N) and ptA).
Chiba-shi, Japan I I
Conditional probabilities p(NIn), ptA In). p(N a), and PtA a) are those
of true negative, false positive, false negative, and true positive diag-
REFERENCES noses, respectively. We can calculate the information content from
these relations. For example. TABLE IV shows the decision matrix of
1. Suzuki R: Congenital Dislocation of Hip-joints and Its Func- =
Doctor 1: p(n) 174/200 = 0.87. pta) = 26/200 = 0.13, P(N) = 0.905,
tional Therapy. (Japanese) Tokyo, Nankodo, 1970 ptA) = 0.095, p(Nln) = 174/174 = 1.0, p{Aln) = 0.0. p{Nla) = 0.269,
2. Lusted LB: Introduction to Medical Decision Making. Thomas,
and p(Ala) = 0.731.
Springfield, III., 1968
The information content can be calculated using equation (A-1) as
3. Metz CE, GoodenoughOJ, Rossmann K: Evaluationof receiver
operating characteristic curve data in terms of information theory, with follows
applications in radiology. Radiology 109:297-303, Nov 1973
4. Onoe M, Takagi M, Masumoto T, et al: Simple input/output
equipment for digital processing of pictorial information. (Japanese) 1= 1.0 x 0.87 x log2(1.0/0.905) + 0.0 x 0.87
J Inst Indust Sci Univ. of Tokyo, 24:127-135, Apr 1972 X log, (0.0/0.095)
5. Kruger RP, Townes TR, Hall DL, et al: Automated radiographic + 0.269 x 0.13 x log2(0.269/0.905)
diagnosis via feature extraction and classification of cardiac size and + 0.731 x 0.13 x log2(0.731/0.095)
shape descriptors. IEEE Trans Sio-Med Eng BME-19:174-186. May = 0.344 (bits)
1972
6. Chow CK, Kaneko T: Automatic boundary detection of the
left ventricle from cineangiograms. Comput Biomed Res 5:388-410, The information content of the input side. H(x),is calculated using p{n)
Aug 1972 and pta) with equation (A-2) as follows
7. Ballard D. Sklansky J: Tumor detection in radiographs. Comput
Biomed Res 6:299-321, Aug 1973
8. Shikano K, Toriwaki J, Fukumura T: Wave propagation H(x) = -0.87 x log2(0.87) - 0.13 x log2(0.13)
method-a method for conversion of grey pictures into line figures. = 0.557 (bits)
Trans Inst Electronics Comm Engnrs Japan 55:668-675. Oct 1972
9. Negoro T, Suenaga Y. Toriwaki J, et al: Ribs identification in
chest photofluorographic films. (Japanase) Technical report of the Thus, I/H(x) = 0.618.

You might also like