Well-defined image can assist user to identify region of interest during segmentation. However, complex medical image is usually characterized by poor tissue contrast and low background luminance. The contrast improvement can lift image visual quality, but the fundamental contrast enhancement methods often overlook the sudden jump problem. In this work, the proposed bihistogram Bezier curve contrast enhancement introduces the concept of “adequate contrast enhancement” to overcome sudden jump problem in knee magnetic resonance image. Since every image produces its own intensity distribution, the adequate contrast enhancement checks on the image’s maximum intensity distortion and uses intensity discrepancy reduction to generate Bezier transform curve. The proposed method improves tissue contrast and preserves pertinent knee features without compromising natural image appearance. Besides, statistical results from Fisher’s Least Significant Difference test and the Duncan test have consistently indicated that the proposed method outperforms fundamental contrast enhancement methods to exalt image visual quality. As the study is limited to relatively small image database, future works will include a larger dataset with osteoarthritic images to assess the clinical effectiveness of the proposed method to facilitate the image inspection.
Magnetic resonance (MR) imaging allows direct visualization of knee cartilage and quantitative measurement on cartilage to monitor osteoarthritis (OA) progression [
There have been growing interests in interactive segmentation recently [
The visual quality of knee MR image can be deteriorated by several phenomena: insignificant tissue contrast, complex knee joint structure, and low background luminance. Definitive tissue contrast facilitates the image interpretation process. Unfortunately, mediocre tissue contrast has been observed in knee MR image. For example, tissue contrast between cartilage, surrounding muscle tissues, and image background only differ slightly. As a result, poor contrast difference observed in the image can easily lead to inter- and intraobserver ambiguity during inspection. Besides, human knee structure is anatomically complex. The knee compartment is filled with various types of cartilages, knee bones, muscle, fat tissue, and ligaments; attempting to interpret MR image equipped with poor tissue contrast is laborious. Further, MR image of knee is characterized by large numbers of low intensity background pixels. Dark background hardens any effective identification of pertinent image features. Eventually, above mentioned phenomena give rise to the prominence of a tissue contrast enhancement model to elevate image visual quality.
Contrast enhancement is defined as a remapping process to transform the image’s intensity distribution so the image intensity range can be fully exploited [
Direct contrast enhancement through traditional histogram equalization [
Fundamental contrast enhancement betterments are developed based on the idea of histogram partitioning. The gist of these improved models concentrates on preserving mean brightness of original image using different intensity threshold values, that is, mean and median. As such, one of the earliest bihistogram equalization methods, brightness preserving bihistogram equalization (BBHE) [
In 1999, Wang et al. [
Chen and Ramli [
In this paper, a new contrast enhancement model known as bihistogram Bezier curve contrast enhancement (BBCCE) is proposed. The intention of BBCCE is to improve image visual quality through appropriate contrast enhancement without compromising the conservation of salient knee features. As such, transformation of intensity distribution is performed using remapping process based on Bezier transform curve instead of conventional cumulative density function. Besides, the important property of bihistogram equalization is retained by partitioning the histogram into two subhistograms to curtail predominance of low intensity background pixels.
Twenty healthy baseline (Data set: 0.C.2) dual-echo steady-state (DESS) knee MR images with water excitation (we) [
Conventional transform curve is derived from cumulative density function. The objective is to stretch the original intensity distribution using transform curve to cover full dynamic range of the image. Thus, image’s contrast can be modified. However, cumulative density function based remapping process does not always yield the desired effect because distortion by dominant intensity levels (as indicated by arrow in Figure
Flow of BBCCE computation using knee MR image. (a) Original MR image. (b) Upper diagram shows histogram of original MR image where black arrow indicates dominant background intensities and lower diagrams show the decomposition of histogram into lower and upper subhistograms in BBCCE. (c) Cumulative density function of original MR image where black arrows indicate large intensity distortion which contributes to sudden jump issue. (d) Intensity discrepency curve duduced from cumulative density function where upward and downward red arrows indicate global maximum and global minimum in lower histogram while upward black arrow indicates global maximum in upper histogram. Leftward black arrow defines the boundary for global extremum in intensity discrepency curve. (e) Bezier transform curve generated using control points duduced from intensity discrepency curve. (f) BBCCE enhanced MR image.
The term “adequate” implies dynamic adjustment of transformed intensity values to refine the traditional transform curve. In addition, we must preserve prominent image features and maintain natural visual appearance simultaneously. Hence, the proposed BBCCE uses the largest intensity discrepancy value deduced from intensity difference curve (Figure
MR image is defined with
In this work, two subhistograms are produced after histogram decomposition. The partition is intended to confine any abrupt distribution skew caused by large amounts of background pixels to lower histogram. To compute histogram decomposition, the mean MR image pixel intensity
Histogram of MR image can be derived from the plot of
As shown in (
In proposed method, the solution for this issue is to smooth the conventional cumulative density function based on intensity distortion caused by sudden jump. Noteworthy, the degree of contrast enhancement depends on absolute intensity difference (AID), which is defined as
The intensity difference curves are deduced by computing discrete AIDs for both subhistograms. Notably, intensity difference curve for each MR image is distinct. This novel feature in the proposed BBCCE allows us to consider arbitrary intensity distribution variation for smoothing conventional cumulative density function. Thus, the transform curve is based on two factors: input intensity distribution and resultant intensity fluctuation. Suppose that the intensity difference curve is defined as
Given that the proposed method defines MR image within
Bezier curve is popular in computer graphics and computer-aided application. The parametric curve, representing the special case of
Bezier curve uses Bernstein polynomial
Based on intensity difference curve, Bezier curve of second degree or third degree is likely to be generated. If a pair of global extrema is detected from the curve, Bezier curve of third degree,
The qualitative assessment and statistical analysis is conducted to evaluate the performance of traditional histogram equalization (referred to THE in Section
Given an original image,
Mean brightness difference between original image and resultant image reveals the degree of luminance distortion. The statistical evaluation metric, known as absolute mean brightness error (AMBE), is defined as
Feature similarity index model (FSIM) aims to perform image quality assessment (IQA) [
We considered several factors when assessing visual quality of BBCCE enhanced image such as natural looking, tissue contrast improvement, preservation of knee features, and minimum image artifact provocation. Effect of BBCCE enhancement is illustrated in Figure
Manifestations of contrast enhancement effect on MR image from medial, central, and lateral sides of the knee joint to represent overall enhancement impact on the stack of 2D MR images. (First row from left to right) Original MR knee image from different sides: (a) medial, (b) central, and (c) lateral. (Second row from left to right) BBCCE enhanced MR knee image from different sides: (d) medial, (e) central, and (f) lateral. Knee cartilage is known as (in red arrows with label): (1) patellar cartilage, (2) femoral cartilage, (3) tibial cartilage. Prominent knee features in this image include (in white labeled arrows): (1) femoral sulcus and (2) Intensity variation within cartilage.
Besides, BBCCE is compared to fundamental contrast enhancement techniques with relative to original MR image. Knee joint is a compound joint packed with various skeletal elements. Interpretation of complex MR image could lead to human ambiguity as a result of unclear structural delineation and unpleasant background image illustration. Figure
Central region manifestation of the left patellofemoral joint section using DESSwe MR imaging sequence. Labels in (a) indicate various skeletal elements inside knee joint (P = patellar, T = tibia, F = femur; 1 = femoral sulcus, 2 = Hoffa fat pad, 3 = anterior cruciate ligament (ACL), 4 = oblique popliteus ligament, 5 = posterior cruciate ligament (PCL), 6 = gastrocnemius muscle, 7 = popliteus muscle, 8 = soleus muscle). For comparison purpose, original MR image is remanifested in (b). Enhanced MR images using (c) THE, (d) BBHE, (e) DSIHE, (f) RMSHE, (g) RSIHE, and (h) BBCCE show the contrast enhancement effect relative to (b). The femoral sulcus (white arrow labeled as “1”) and intensity variation within cartilage (unlabeled white arrow) are indicated in images from (b) to (h).
Typically, all contrast enhancement methods improve the tissue contrast, lifting the image’s background luminance. However, apparent image noise amplification is detected in some previous contrast enhancement methods. For example, image noise amplified by traditional histogram equalization (THE) is obvious especially at femur and tibia. Irritating image artifact downgrades the visual quality. Besides, serious noise amplification is detected at popliteus muscle, gastrocnemius muscle, and soleus muscle in images produced by RSIHE and RMSHE. Thus, THE, RSIHE, and RMSHE are largely unsuitable for medical image contrast enhancement.
Preservation of salient information describing knee cartilage is imperative. Precise structural delineation through adequate tissue contrast refinement can maintain small but important feature details. For instance, BBCCE improves the cartilage contrast as well as conserves paramount details like femoral sulcus (white arrow labeled as “1”) and intensity variation within cartilage (unlabeled white arrow) simultaneously. Previous contrast enhancement methods, on the other hand, over-enhance the cartilage. Hence, intensity variation is destroyed and patellar cartilage is seen to “combine” with femoral cartilage.
All tests are performed using SPSS (version 21). In this study, the hypothesis that BBCCE could outperform other contrast enhancement methods is tested and verified. Table
Mean EME, AMBE, and FSIM values computed from THE, BBHE, DSIHE, RMSHE, RSIHE, and BBCCE by taking original image’s EME (32.38) and AMBE (0.00) as reference. FSIM ranged from 0 to 1, where 1 indicates the best enhanced image quality.
Methods | Mean EME |
95% confidence interval of the difference | Mean AMBE |
95% confidence interval of the difference | Mean FSIM |
95% confidence interval of the difference | |||
---|---|---|---|---|---|---|---|---|---|
Lower | Upper | Lower | Upper | Lower | Upper | ||||
THE |
|
46.66 | 48.58 |
|
75.05 | 79.73 |
|
0.71 | 0.74 |
BBHE |
|
44.79 | 46.60 |
|
26.13 | 28.57 |
|
0.78 | 0.81 |
DSIHE |
|
46.16 | 47.86 |
|
34.71 | 36.90 |
|
0.76 | 0.79 |
RMSHE |
|
42.48 | 44.17 |
|
13.42 | 14.63 |
|
0.82 | 0.85 |
RSIHE |
|
53.86 | 55.62 |
|
16.84 | 17.96 |
|
0.81 | 0.83 |
BBCCE |
|
40.94 | 41.93 |
|
14.13 | 15.51 |
|
0.91 | 0.93 |
Table
The one-way ANOVA computed by using different contrast enhancement methods in EME, AMBE, and FSIM.
Sum of squares | df | Mean square |
|
|
|
---|---|---|---|---|---|
EME | |||||
Methods | 2111.80 | 5 | 422.36 | 132.72 | 0.00* |
Errors | 362.80 | 114 | 3.18 | ||
Total |
|
|
|||
FSIM | |||||
Methods | 58467.66 | 5 | 11693.53 | 1652.39 | 0.00* |
Errors | 806.75 | 114 | 7.08 | ||
Total |
|
|
|||
AMBE | |||||
Methods | 0.42 | 5 | 0.09 | 124.12 | 0.00* |
Errors | 0.08 | 114 | 0.00 | ||
Total |
|
|
Fisher’s least significant difference test (Table
Categorization of different methods using Fisher’s Least Significance Difference (LSD) for EME.
( |
( |
Mean difference ( |
Std. error |
|
95% confidence interval | |
---|---|---|---|---|---|---|
Lower bound | Upper bound | |||||
THE | BBHE | 1.92* | 0.56 | 0.00 | 0.81 | 3.04 |
DSIHE | 0.61 | 0.56 | 0.28 | −0.51 | 1.73 | |
RMSHE | 4.29* | 0.56 | 0.00 | 3.17 | 5.41 | |
RSIHE | −7.12* | 0.56 | 0.00 | −8.23 | −6.00 | |
BBCCE | 6.18* | 0.56 | 0.00 | 5.07 | 7.30 | |
|
||||||
BBHE | THE | −1.92* | 0.56 | 0.00 | −3.04 | −0.81 |
DSIHE | −1.31* | 0.56 | 0.02 | −2.43 | −0.20 | |
RMSHE | 2.37* | 0.56 | 0.00 | 1.25 | 3.48 | |
RSIHE | −9.04* | 0.56 | 0.00 | −10.16 | −7.92 | |
BBCCE | 4.26* | 0.56 | 0.00 | 3.14 | 5.38 | |
|
||||||
DSIHE | THE | −0.61 | 0.56 | 0.28 | −1.73 | 0.51 |
BBHE | 1.31* | 0.56 | 0.02 | 0.20 | 2.43 | |
RMSHE | 3.68* | 0.56 | 0.00 | 2.56 | 4.80 | |
RSIHE | −7.73* | 0.56 | 0.00 | −8.84 | −6.61 | |
BBCCE | 5.57* | 0.56 | 0.00 | 4.46 | 6.70 | |
|
||||||
RMSHE | THE | −4.29* | 0.56 | 0.00 | −5.41 | −3.17 |
BBHE | −2.37* | 0.56 | 0.00 | −3.48 | −1.25 | |
DSIHE | −3.68* | 0.56 | 0.00 | −4.80 | −2.56 | |
RSIHE | −11.41* | 0.56 | 0.00 | −12.53 | −10.29 | |
BBCCE | 1.89* | 0.56 | 0.00 | 0.78 | 3.01 | |
|
||||||
RSIHE | THE | 7.12* | 0.56 | 0.00 | 6.00 | 8.23 |
BBHE | 9.04* | 0.56 | 0.00 | 7.92 | 10.16 | |
DSIHE | 7.73* | 0.56 | 0.00 | 6.61 | 8.84 | |
RMSHE | 11.41* | 0.56 | 0.00 | 10.29 | 12.53 | |
BBCCE | 13.30* | 0.56 | 0.00 | 12.18 | 14.41 | |
|
||||||
BBCCE | THE | −6.18* | 0.56 | 0.00 | −7.30 | −5.07 |
BBHE | −4.26* | 0.56 | 0.00 | −5.38 | −3.14 | |
DSIHE | −5.57* | 0.56 | 0.00 | −6.69 | −4.46 | |
RMSHE | −1.89* | 0.56 | 0.00 | −3.01 | −0.78 | |
RSIHE | −13.30* | 0.56 | 0.00 | −14.42 | −12.18 |
Categorization of contrast enhancement methods into homogenous subset using the Duncan test for EME.
Method |
|
Subset for alpha = 0.05 | ||||
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
BBCCE | 20 | 41.44 | ||||
RMSHE | 20 | 43.33 | ||||
BBHE | 20 | 45.70 | ||||
DSIHE | 20 | 47.01 | ||||
THE | 20 | 47.62 | ||||
RSIHE | 20 | 54.74 | ||||
Sig. | 1.00 | 1.00 | 1.00 | 0.28 | 1.00 |
Means for groups in homogeneous subsets are displayed.
Harmonic Mean Sample Size = 20.000.
Fisher’s least significant difference test (Table
Categorization of different methods using Fisher’s Least Significance Difference (LSD) for AMBE.
( |
( |
Mean difference ( |
Std. error |
|
95% confidence interval | |
---|---|---|---|---|---|---|
Lower bound | Upper bound | |||||
THE | BBHE | 50.04* | 0.84 | 0.00 | 48.38 | 51.71 |
DSIHE | 41.59* | 0.84 | 0.00 | 39.92 | 43.25 | |
RMSHE | 63.37* | 0.84 | 0.00 | 61.70 | 65.03 | |
RSIHE | 60.00* | 0.84 | 0.00 | 58.32 | 61.66 | |
BBCCE | 62.57* | 0.84 | 0.00 | 60.91 | 64.24 | |
|
||||||
BBHE | THE | −50.04* | 0.84 | 0.00 | −51.71 | −48.38 |
DSIHE | −8.45* | 0.84 | 0.00 | −10.12 | −6.79 | |
RMSHE | 13.33* | 0.84 | 0.00 | 11.66 | 14.99 | |
RSIHE | 9.95* | 0.84 | 0.00 | 8.28 | 11.62 | |
BBCCE | 12.53* | 0.84 | 0.00 | 10.86 | 14.20 | |
|
||||||
DSIHE | THE | −41.59* | 0.84 | 0.00 | −43.25 | −39.92 |
BBHE | 8.45* | 0.84 | 0.00 | 6.79 | 10.12 | |
RMSHE | 21.78* | 0.84 | 0.00 | 20.11 | 23.45 | |
RSIHE | 18.40* | 0.84 | 0.00 | 16.74 | 20.07 | |
BBCCE | 20.98* | 0.84 | 0.00 | 19.32 | 22.65 | |
|
||||||
RMSHE | THE | −63.37* | 0.84 | 0.00 | −65.03 | −61.70 |
BBHE | −13.33* | 0.84 | 0.00 | −15.00 | −11.66 | |
DSIHE | −21.78* | 0.84 | 0.00 | −23.45 | −20.11 | |
RSIHE | −3.38* | 0.84 | 0.00 | −5.04 | −1.71 | |
BBCCE | −0.79 | 0.84 | 0.35 | −2.46 | 0.87 | |
|
||||||
RSIHE | THE | −60.00* | 0.84 | 0.00 | −61.66 | −58.32 |
BBHE | −9.95* | 0.84 | 0.00 | −11.62 | −8.28 | |
DSIHE | −18.40* | 0.84 | 0.00 | −20.07 | −16.74 | |
RMSHE | 3.38* | 0.84 | 0.00 | 1.71 | 5.04 | |
BBCCE | 2.58* | 0.84 | 0.00 | 0.92 | 4.25 | |
|
||||||
BBCCE | THE | −62.57* | 0.84 | 0.00 | −64.24 | −60.91 |
BBHE | −12.53* | 0.84 | 0.00 | −14.20 | −10.87 | |
DSIHE | −20.98* | 0.84 | 0.00 | −22.65 | −19.32 | |
RMSHE | 0.79 | 0.84 | 0.35 | −0.87 | 2.46 | |
RSIHE | −2.58* | 0.84 | 0.00 | −4.25 | −0.92 |
Categorization of contrast enhancement methods into homogenous subset using the Duncan test for AMBE.
Method |
|
Subset for alpha = 0.05 | ||||
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
RMSHE | 20 | 14.02 | ||||
BBCCE | 20 | 14.82 | ||||
RSIHE | 20 | 17.40 | ||||
BBHE | 20 | 27.35 | ||||
DSIHE | 20 | 35.80 | ||||
THE | 20 | 77.39 | ||||
Sig. | 0.35 | 1.00 | 1.00 | 1.00 | 1.00 |
Means for groups in homogeneous subsets are displayed.
Harmonic Mean Sample Size = 20.000.
Fisher’s least significant difference test (Table
Categorization of different methods using Fisher’s Least Significance Difference (LSD) for FSIM.
( |
( |
Mean difference ( |
Std. error |
|
95% confidence interval | |
---|---|---|---|---|---|---|
Lower bound | Upper bound | |||||
THE | BBHE | −0.07* | 0.01 | 0.00 | −0.09 | −0.05 |
DSIHE | −0.05* | 0.01 | 0.00 | −0.06 | −0.03 | |
RMSHE | −0.11* | 0.01 | 0.00 | −0.12 | −0.09 | |
RSIHE | −0.10* | 0.01 | 0.00 | −0.11 | −0.08 | |
BBCCE | −0.19* | 0.01 | 0.00 | −0.21 | −0.18 | |
|
||||||
BBHE | THE | 0.07* | 0.01 | 0.00 | 0.05 | 0.09 |
DSIHE | 0.02* | 0.01 | 0.01 | 0.00 | 0.04 | |
RMSHE | −0.04* | 0.01 | 0.00 | −0.05 | −0.02 | |
RSIHE | −0.03* | 0.01 | 0.00 | −0.04 | −0.01 | |
BBCCE | −0.12* | 0.01 | 0.00 | −0.14 | −0.11 | |
|
||||||
DSIHE | THE | 0.05* | 0.01 | 0.00 | 0.03 | 0.06 |
BBHE | −0.02* | 0.01 | 0.01 | −0.04 | −0.00 | |
RMSHE | −0.06* | 0.01 | 0.00 | −0.08 | −0.04 | |
RSIHE | −0.05* | 0.01 | 0.00 | −0.06 | −0.03 | |
BBCCE | −0.15* | 0.01 | 0.00 | −0.16 | −0.13 | |
|
||||||
RMSHE | THE | 0.11* | 0.01 | 0.00 | 0.09 | 0.12 |
BBHE | 0.03* | 0.01 | 0.00 | 0.02 | 0.05 | |
DSIHE | 0.06* | 0.01 | 0.00 | 0.04 | 0.08 | |
RSIHE | 0.01 | 0.01 | 0.12 | −0.00 | 0.03 | |
BBCCE | −0.09* | 0.01 | 0.00 | −0.10 | −0.07 | |
|
||||||
RSIHE | THE | 0.09* | 0.01 | 0.00 | 0.08 | 0.11 |
BBHE | 0.03* | 0.01 | 0.00 | 0.01 | 0.04 | |
DSIHE | 0.05* | 0.01 | 0.00 | 0.03 | 0.06 | |
RMSHE | −0.01 | 0.01 | 0.12 | −0.03 | 0.00 | |
BBCCE | −0.10* | 0.01 | 0.00 | −0.12 | −0.08 | |
|
||||||
BBCCE | THE | 0.19* | 0.01 | 0.00 | 0.18 | 0.21 |
BBHE | 0.12* | 0.01 | 0.00 | 0.11 | 0.14 | |
DSIHE | 0.15* | 0.01 | 0.00 | 0.13 | 0.16 | |
RMSHE | 0.09* | 0.01 | 0.00 | 0.07 | 0.10 | |
RSIHE | 0.10* | 0.01 | 0.00 | 0.08 | 0.12 |
Categorization of contrast enhancement methods into homogenous subset using Duncan’s test for FSIM.
Method |
|
Subset for alpha = 0.05 | ||||
---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | ||
THE | 20 | 0.72 | ||||
DSIHE | 20 | 0.77 | ||||
BBHE | 20 | 0.79 | ||||
RSIHE | 20 | 0.82 | ||||
RMSHE | 20 | 0.83 | ||||
BBCCE | 20 | 0.92 | ||||
Sig. | 1.00 | 1.00 | 1.00 | 0.12 | 1.00 |
Means for groups in homogeneous subsets are displayed.
Harmonic Mean Sample Size = 20.000.
The performance of contrast enhancement methods (THE, BBHE, DSIHE, RMSHE, RSIHE, and BBCCE) is ranked according to the results computed from EME, AMBE, and FSIM in Table
Ranking of methods in terms of enhancement degree (EME), image quality (FSIM), and mean intensity distortion (AMBE). The methods ranking is computed according to Fisher’s Least Significance Difference (LSD) and the Duncan test.
Rank | EME | FSIM | AMBE |
---|---|---|---|
1 |
|
|
RMSHE, |
2 | RMSHE | RMSHE, RSIHE | |
3 | BBHE | RSIHE | |
4 | THE, DSIHE | BBHE | BBHE |
5 | DSIHE | DSIHE | |
6 | RSIHE | THE | THE |
In this paper, bihistogram Bezier curve contrast enhancement (BBCCE) is presented. The objective of BBCCE is to assist radiologist to interpret MR image prior to performing knee cartilage segmentation. However, MR image possesses poor tissue contrast, and conventional contrast enhancement methods have failed to address this issue. As such, sudden jump in conventional transform curve causes the image to be over-enhanced, and deteriorates the image visual quality. Therefore, it is believed that the adequate contrast enhancement is the most appropriate solution as in the case of MR image. To achieve the objective, the Bezier transform curve based on intensity discrepancy value and intensity difference curve is deduced. The quantitative results show that BBCCE excels in terms of tissue improvement, minimal mean intensity distortion, and image quality. The results are in-line with qualitative results which show that BBCCE could preserve important knee features and better delineate the knee structure without provoking much image noise. In the future, the mean opinion survey and record image evaluation time to assess the effectiveness of BBCCE will be performed in assisting radiologists to interpret knee MR image.
The authors declare that there is no conflict of interests regarding the publication of this paper.
The authors gratefully acknowledge the Ainuddin Wahid scholarship provided by School of Graduate Studies, Universiti Teknologi, Malaysia, and university research grant provided by Research Management Centre and sponsored by Ministry of Higher Education, Malaysia. Vot: Q.J130000.2545.04H41, GUP Universiti Teknologi Malaysia, Johor Bahru, Malaysia.