The low quality of diagnostic Ultrasound (US) images makes structural identification challenging. It is crucial to enhance these images before application in Computer-Aided Detection (CAD). This study aims to ease breast lesion identification in US images by introducing a novel echo-texture-based technique. We used the basic notion that the unique acoustic impedance of distinct structures results in a textural discrepancy in the US image. Initially, an Acoustic Impedance Index was calculated using the local grayscale variation in the image. Then, we used this parameter to design an adaptive filter that generates an Enhanced Texture Image. The resultant images had a significant perception difference between the lesion and its surrounding. Compared to other enhancement techniques, the No Reference distortion metric had the best values for the proposed method. The enhanced image segmented using Active contour and superpixel-based techniques gives the most accurate lesion boundary. The Dice, Jaccard, and BF score were used to quantify the similarity between the masked image generated after segmentation and outlined by the expert. With the exceptionally reduced probability of false alarm, the proposed technique promises to improve the CAD of breast lesions.
Research Article
A Novel Framework for Ultrasound Image Enhancement to assist segmentation in Computer-Aided Detection of Breast Lesion
https://doi.org/10.21203/rs.3.rs-1552793/v1
This work is licensed under a CC BY 4.0 License
You are reading this latest preprint version
Acoustic Impedance
Breast Cancer
Computer-Aided Detection
Enhancement
Segmentation
Ultrasound
The imminent demand for Computer-Aided Detection (CAD) can be conceived from the escalating cases of breast cancer around the globe [1]. For breast cancer, Ultrasound imaging is the only diagnostic imaging modality that is safe and economical [2]. So, it is the most suitable platform to design the CAD system. Skillful structure identification is crucial for the best outcome by the CAD [3]. Unfortunately, the quality of Breast Ultrasound (BUS) is not adequate for any structural identification [4]. A trivial judgment by CAD due to the low-quality image may lead to a fatal consequence. So, the images need to be processed for the best outcomes.
The researchers have explored a diverse framework to date for image enhancement. Initially, image histogram modifications were practiced for image enhancement. Like Histogram Equalization (HE), the image contrast is improved by stretching the histogram. Later different studies introduced different upgrades of HE to enhance the image contrast. Contrast Limited Adaptive Histogram Equalization (CLAHE) is the most practiced update of HE [5]. The images enhanced by these techniques show an increased contrast but eventually turn into over-enhanced images. Another framework Gamma Correction has been used to enhance and compress the image simultaneously. It is a pixel-by-pixel operation where each pixel is raised to the power of some Gamma. The choice of Gamma has always been the prime target for modification by different studies. Using weighted cumulative distribution function for Gamma in Adaptive Gamma Correction with Weighting Distribution (AGCWD) by S. C. Huang et al. in (2013) is still most practiced enhancement method. But it is not often preferred for medical images that primarily have a dark texture. Retinex adopts a unique technique to enhance images that are based on human visual perception introduced by E. H. Land [7]. Here the input image is processed to obtain an acceptable approximation of the reflectance image. But it has limited application in grayscale images, which was improved in [7] as Multi-Scale Retinex (MSR). Where he set three different scales for Retinex, and the scaled images were later averaged to obtain the output image. This improvement was very efficient but still not good enough to exercise on medical images. Some studies consider image enhancement as a task to reduce the speckle noises in the ultrasound images while preserving the edge. That also improves the results of post-processing steps like segmentation which will see later in the result section. Prerna et al. in applied a combination of clustering for wavelet sub-bands (CWS) of the image to reduce the inherent speckle. The study accomplished remarkable results in terms of quality metrics. But applying an edge detection algorithm before denoising and adding later to enhance edges can lead to spurious segmentation. A recent study [8] proposed an enhancement technique by generating a standardized histogram of echotextures for histogram matching (SHEHM). But due to the lack of comparative analysis with other studies, its superiority for practice in medical diagnosis remains unexplored.
After having an enhanced image, the next step in a CAD system is lesion boundary detection employing any segmentation techniques. There is a stack of segmentation methods, but some are eminently suitable for lesion boundary detection. Researchers have been using Active Contour (AC) based segmentation very frequently for medical image segmentation [9]–[12]. In this technique, an initial contour starts evolving in the presence of different forces. The forces drive the contour towards a higher gradient while maintaining contour continuity. Another diagnostic image segmentation method uses the image to divide into groups of superpixels and then segment and systematically link the divided regions [13]–[16]. The efficacy of this technique depends on how well the superpixel division adheres to the object boundary. So, [17] designed a Multi-Scale Superpixel (MSS) technique for object boundary segmentation which shows promising results. Both are efficient and accurate enough to be adopted for segmentation in a CAD system. The credibility of these segmentation techniques is fragile due to the quality of the images.
All fore mentioned studies for image enhancement, one way or another, attempt to modify the contrast of the entire image. However, the enhancement of BUS for CAD application must strengthen structural identification. A significant research work conducted in [18] concluded that it is possible to characterize tissue from its acoustic impedance. Later the research by [19] concluded the difference in acoustic impedance can be adopted to discriminate between cancerous and non-cancerous cells. Rather many studies have used textural details for the classification of breast lesions. But very few have used this textural feature to enhance the ultrasound image because this can help structural-based subjective enhancement. So here, we propose a novel image enhancement technique for BUS images. This study leverages the anomalies in BUS textures, which corresponds to acoustic impedance anomalies. Every underlying structure in the breast has its unique acoustic impedance and can be used to characterize the lesion in BUS. So, we calculated a local textural-based metric that can be considered the pixel's Acoustic Impedance Index (AII). Then the complete image was adaptively filtered using the AII to generate the Enhanced Textural Image (ETI). The ETI was further processed using AGCWD to decrease the distortions. The contribution of this study is as follows.
-
Based on acoustic impedance, we have developed an enhancement technique capable of structure (characterized by acoustic impedance) based enhancement.
-
The developed enhancement technique supports the structural identification process by isolating the lesion from its surrounding.
-
The output image of the proposed technique has the most negligible distortions (like speckle noise or any other noise)
-
Our technique also supports segmentation using AC or MSS to detect lesion boundaries and has negligible chances of false alarm.
This study provides a detailed comparison of enhancement techniques in practice with the proposed method using quality metrics. We have also analyzed the segmentation using recent techniques and discussed the influence of the proposed methodology on CAD. The proposed method significantly improves the active contour and superpixel-based lesion boundary detection.
When ultrasonic radiation is flashed in the vicinity of the underlying mass, it bounces back depending upon the absorbing tendency of the mass. The reflection of the ultrasonic radiation at any surface is characterized using difference in acoustic impedance. The receptor sensors use these reflected waves to generate a perception of the underlying structure. Since the acoustic reflection is a function of acoustic impedance, the textures in the output image are also a function of acoustic impedance. So, it is more reasonable to use the textures in the BUS to calculate an AII of the corresponding pixel. Moreover, using this AII as a primary parameter to enhance the image can support a subjective enhancement. The complete methodology is divided into steps and given in Fig. 1.
Acoustic Impedance Index
Assuming one pixel of the BUS corresponds to a unit volume of a mass of the exposure region. The intensity of the acoustic wave reflected by this mass will reflect in the grayscale value of the corresponding pixel \(i(m,n)\). This phenomenon can be mathematically formulated using acoustic impedance, where the intensity of the reflection is directly proportional to the difference in acoustic impedance. Here we have assumed that the variation of acoustic impedance within a similar volume of masses is negligible. We introduced an AII \({z}_{(m,n)},\) which parametrizes the intensity variation due to acoustic impedance. It is an estimation of the impedance, and its calculation is given in (1).
$${z}_{(m,n)}=b\times \left(\sum _{r,s}\left(i\right(m\pm r,n\pm s)⁄\alpha )\right)$$
1
Here \({z}_{(m,n)}\) represents the acoustic impedance of pixel at coordinate \(m,n (m\in N and n\in N). r and s\)controls the extent of involvement of neighbor pixels in the prediction of \(z\). It is recommended to use smaller values for \(r and s\) because larger values will increase the interference of the neighboring pixels in the estimation. Moreover, point estimation \((r=s=0)\) will make the estimation prone to noise. The multiplier b depends upon the choice of\(r\) and \(s\). It is the inverse of the product of the range of \(r and s\). The normalization constant α assures that \(\left|z\right|\le 1\), and for a grayscale image, it is 255.
It is to be noted that the impedance values obtained or used in this study were not the actual acoustic impedance of the relevant mass. These values are drawn concerning the output image of the ultrasound transducer. We have considered a pixel of value 255 as a perfect reflector and vice versa.
Enhanced Textural Image
The variation in intensity corresponds to variation in acoustic impedance in the BUS. So, we utilized the difference between the adjacent pixel’s impedance parameter \(\left(dz\right)\) to design an adaptive filter \(F\) as given in (2).
The value of G is given in (3), and the calculation for \(\sigma\) is given in (4).
$$G=\left[\begin{array}{ccc}2& 1& 2\\ 1& 0& 1\\ 2& 1& 2\end{array}\right]$$
3
$$\sigma =\left|{\sigma }_{0}+dz\right|$$
4
$$dz\left(m,n\right)=\varDelta z\left(m,n\right)$$
5
Here the \(\varDelta\) is the difference operator, which calculates the difference between the current and adjacent values of \(z\). The initial value of \(\sigma\) is represented as \({\sigma }_{0}\), which is updated within filtering. The initial value of the impedance parameter z is represented as \({z}_{0}\). The calculation of ETI is very sensitive to the initialization of parameters. Random values of \({z}_{0}\) can cause false mapping of impedance variation. So, in this study, we define \({z}_{0}\) as in (6).
$${z}_{0}={b}_{0}\left({\sum }_{m,n}\left(i\right(m,n)⁄\alpha )\right)$$
6
Where α is similar to (2), used for calculation of \(z\). The influence of pixels and region of the image for prediction of initial impedance is controlled by \(m and n\). Which is similar to \(r and s\), and the calculation of constant \({b}_{0}\) is similar to calculation of\(b\) in (1). We preferred the prediction to involve global intensities instead of a local intensity, for an unbiased initialization. In the beginning, \({\sigma }_{0}\)was taken equal to the initial impedance parameter \({z}_{0}\). At each step of filtering \({\sigma }_{0}\)is updated with the last value of \(\sigma\), and the next \(\sigma\) is calculated by adding this \({\sigma }_{0}\) with the difference of acoustic impedance \(dz\). From the designed filter \(F\), ETI can be calculated by (7).
$$ETI=\left(F*I\text{'}\right)$$
7
Where \({I}^{{\prime }}\)is \(I\)after removing the padded pixels, and * is the convolutional operator. The ETI obtained still has some unsharp boundaries and the presence of noise. It can lead to false-positive in the prediction of cancer in CAD. Since ETI has many bright pixels, we have used AGCWD [19] to enhance the ETI further. Figure 2(c) shows the image obtained after processing the ETI with the AGCWD.
|
Algorithm |
|---|
|
Input image \(I(m,n)\) |
|
Calculate average impedance \({z}_{0}\) using Eq. 6, |
|
Use this \({z}_{0}\) as the initial value of \({\sigma }_{0}\) and dz = 0 |
|
For1\(m=1 to M+1\) |
|
For2\(n=1 to N+1\) |
|
Calculate the AII \({(z}_{(m,n)})\) from Eq. 1 |
|
Use Eqs. 2, 3 & 4 to calculate F |
|
Update the value of \({\sigma }_{0}\) |
|
Calculate \(dz\) with equations 5 |
|
Increment\(n\) |
|
End For2 |
|
Increment \(m\) and start with\(n=1\) |
|
End For1 |
|
Use Eq. 7 to calculate ETI |
|
Apply AGCWD to ETI to obtain\({I}_{out}\) |
|
Return\({I}_{out}\) |
The proposed technique was applied to two BUS datasets. The combined data set has 943 BUS images, 597 BUS with benign lesions, 263 with malignant lesions, and 133 non-lesion images. After closely analyzing the BUS, masked images were generated by the experts. Both datasets were analysed by the experts. Complete details about the datasets are available in [20], [21].
Visual Assessment
To represent the efficiency of the proposed technique, we have conducted a comparative analysis with other prevailing image enhancement techniques. There are two categories of quality metrics based on the involvement of reference images, namely Full Reference and No-Reference quality metrics. A Full reference quality metric determines the quality of the image after comparison with a reference image that is supposed to be of high quality. While the no-reference quality metric does not need a reference image, it takes the image and applies an algorithm to predict its quality. This quality assessment technique is also known as Objective blind image quality assessment. From previous segmentation results, it is clear that no high-quality reference image exists. In the absence of a high-quality reference image, the non-reference quality assessment metric was used.
The Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) metric defined in [22] is statistically better than many Full reference quality metrics. Another quality metric is Natural Image Quality Evaluator (NIQE) [23], which works on the natural scene static model and some static features to measure the deviation in statistical regularities from the natural image. Without any human evaluated distortion image, it is entirely blind. Along with NIQE and BRISQUE, we have also adopted the Entropy of the enhanced images.
Fig 3 has the pictorial presentation of the enhanced images, and Table 1 presents the average values of the quality metrics. All the metrics attend a lower value for a less distorted image and vice versa. BRISQUE, NIQE, and Entropy values for the original image (43.01, 7.744, and 7.083, respectively) set the standard for other enhancement images. Any improvement will be reflected by the lower metric values in the table. AGCWD has amplified the distortions already present in image 3(b). AGCWD has metrics values in Table 1 similar to the original image. The BRISQUE value for AGCWD is 43.45 slightly more significant than the original image. The CLAHE enhanced images show a minimal improvement in the metric values compared to the original image. In contrast, the techniques introduced by Prerna et al. and Roy et al. show far better performance than the CLAHE and AGCWD in Table 1. The reduction in distortion can be easily observed at the bottom of Fig 3 (j) and (k). The values achieved by the proposed technique for BRISQUE, NIQE, and Entropy are 38.87, 5.57, and 4.01, respectively. The proposed method has the least values for all metrics. It shows that the images enhanced by the proposed technique have the least distortion.
Segmentation Assessment
Here we have analyzed the result of various enhancement techniques segmented using AC and the MSS. After segmentation of the BUS image, we have generated the mask for corresponding segmentation for comparison. The mask images generated by the expert were used to find the similarity metric value. For quantitative evaluation, we used the DICE similarity metric[24], Jaccard similarity [25], and BF score [26]. This study took the recent enhancement techniques to compare efficacy in lesion segmentation. Fig 4 shows the masked image of different techniques segmented using the AC method. Fig 4 (a) & (b) are the original BUS image with lesions of different sizes and shapes. The images enhanced using CWS have very much segmented the lesion boundary where the intensity change is significant. But for the boundary regions where the intensity variation is minimal, it fails to adhere to the lesion boundary. The average similarity metric values encountered across the datasets have been provided in Table 2, and CWS has the most negligible value of similarity metrics. For SHEHM, the similarity metric values in Table 2 are better than CWS but still not adequate. The masked version of the image enhanced using the SHEHM in Fig 4 (c) & (g) shows the penetration of the segmented regions in the surrounding. The masked image of the proposed technique completely outruns both methods for boundary detection in Fig 4 (d) & (h). The enhanced image’s segmentation adheres closest to the true lesion boundary. The values of Dice, Jaccard, and BF score for the mask of the image enhanced using the proposed technique are 0.8179, 0.69, and 0.4818, respectively.
Fig 5 shows the edge detection using the MSS and simple superpixel applied to the output image of the proposed technique. They have lesser spikes at the boundary than the masked image generated after AC. All of the three techniques for segmentation adhere well over the lesion boundary of the image enhanced using the proposed technique. Edges generated by superpixel in Fig 5 (c) and (f) have better boundaries than the edges of active contour-based segmentation in Fig 4 (d) and (h). Moreover, the similarity metric values for the superpixel-based edge detection are far better than others.
False Positive assessment
For a reliable CAD system, it’s crucial to analyse the system slump towards false positives. To verify this, we applied the proposed technique to the non-lesion images. The segmentation of the enhanced image must not adhere to some random intensity anomaly. We have calculated the probability of false-positive across the dataset as an evaluation metric in Table 3. Fig 6 (a) and (b) are two original non-lesion images from the dataset. The enhanced images were segmented using the AC method. The lesion candidates were selected based on standard features used in different studies [27], [28]. The CLAHE and the AGCWD enhanced images have caught up to some intensity anomalies in Fig 3 (b, c, h & j). And from the table, it is clear that both techniques are too prone to false positives. The performance of the de-noising technique is better than the CLAHE and AGCWD. But its segmentation using AC also caught up to a slight variation in BUS. The image enhancement technique proposed by Roy et al. was just better than CWS. However, the performance of all the techniques was competitive which each other but still not adequate for common practice. Fig 6 (l) & (f) the masked image has no traces of any blob-like structure. It means no trace of a lesion in that BUS, and it also proves that the segmentation of the proposed technique is not vulnerable to random grayscale anomalies.
The results stated in the previous section compares values of different performance metric for the proposed method with some practiced technique. The proposed method has been developed based on the relation between acoustic impedance and corresponding textural appearance. This relationship helped us achieve subjective enhancement and successfully isolate the lesion to ease its detection. It can be confirmed from Figs. 2 and 3 that the lesion is separated from the surrounding. The technique further smooths out the inherent distortion and has the least BRISQUE, NIQE, and Entropy values. In contrast, the other methods did not aim to isolate the lesion but performed well in reducing distortion. Like CLAHE has enhanced the image's brightness but has also smoothed out the lesion boundary. For AGCWD, many studies have concluded its sparse efficiency for images with dark textures. The same can be observed in Table 1, where the AGCWD enhanced image has greater values than the original BUS. The denoising techniques like the CWS have reduced the distortion but are still not efficient enough for lesion detection. And same goes for the SHEHM, which can be observed from the mask images generated from CWS and SHEHM. The masked images generated after segmentation using AC breached the surrounding of the lesion. When used for lesion characterization or classification, this can lead to a severe problem. Because of the subjective enhancement achieved by our technique, lesion identification is simplified. Even basic superpixel-based segmentation technique shows promising results. The masked images were approximately similar to the masked images generated by the experts. So, we had much higher values of the similarity metrics in Table 2. One way to judge the efficacy of the proposed technique in identifying different regions in BUS is by the probability of false alarm. The false alarm probability should have been higher if the developed method was prone to just any grayscale variation. But the possibility of false alarm is fairly low compared to others. It can be further reduced if channeled with an advanced segmentation technique or features introduced the in future for lesion identification.
This study implemented an enhancement technique for BUS that can ease the identification of any underlying lesion. It explores acoustic impedance differences across masses to isolate the breast lesion. We calculated an AII for each pixel, considering the impedance dependency of ultrasound image texture. An adaptive filter was designed based on the calculated AII to generate an ETI. This ETI was later used for gamma correction to obtain the final output image. In the visual assessment of the resultant image, it was found that the lesion was isolated from its surrounding. There was a uniform enhancement in regions of similar acoustic impedance, so the proposed technique does not introduce any disturbance within the homogeneous area. The proposed technique observed null chances of false alarms in this study. The segmentation of any existing tumor in BUS proved to be the best enhancement technique among all other prevailing methods. It also increased the efficacy of the prevailing segmentation technique in lesion edge detection. The output image has the least distortion compared to CLAHE, AGCWD, CWS, and SHEHM. It has the highest scores for the No-Reference quality and similarity metrics. Enhancement was homogeneous in impedance-wise localized regions.
- This work has no source of funding to declare
- Authors does not share any conflict of interest and this submission is approved by all authors
- No ethical approval required
- Wang Y et al (2020) “Deeply-Supervised Networks With Threshold Loss for Cancer Detection in Automated Breast Ultrasound,” IEEE Trans. Med. Imaging, vol. 39, no. 4, pp. 866–876, doi: 10.1109/TMI.2019.2936500
- Kolb TM, Lichy J, Newhouse JH (2002) “Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations.,” Radiology, vol. 225, no. 1, pp. 165–175, Oct. doi: 10.1148/radiol.2251011667
- Sudharson S, Kokil P (2021) Computer-aided diagnosis system for the classification of multi-class kidney abnormalities in the noisy ultrasound images. Comput Methods Programs Biomed 205:106071. doi: https://doi.org/10.1016/j.cmpb.2021.106071
- Chen C, Wang Y, Niu J, Liu X, Li Q, Gong X (2021) Domain Knowledge Powered Deep Learning for Breast Cancer Diagnosis Based on Contrast-Enhanced Ultrasound Videos. IEEE Trans Med Imaging 40(9):2439–2451. doi: 10.1109/TMI.2021.3078370
- Reza AM (2004) Realization of the Contrast Limited Adaptive Histogram Equalization (CLAHE) for Real-Time Image Enhancement. J VLSI signal Process Syst signal image video Technol 38(1):35–44. doi: 10.1023/B:VLSI.0000028532.53893.82
- Huang S-C, Cheng F-C, Chiu Y-S (2013) Efficient Contrast Enhancement Using Adaptive Gamma Correction With Weighting Distribution. IEEE Trans Image Process 22(3):1032–1041. doi: 10.1109/TIP.2012.2226047
- Petro AB, Sbert C, Morel J-M (2014) “Multiscale retinex,”Image Process. Line, pp.71–88,
- Roy R, Ghosh S, Ghosh A (2020) Clinical ultrasound image standardization using histogram specification. Comput Biol Med 120:103746. doi: https://doi.org/10.1016/j.compbiomed.2020.103746
- Lotfollahi M, Gity M, Ye JY, Mahlooji Far A (2018) Segmentation of breast ultrasound images based on active contours using neutrosophic theory. J Med Ultrason 45(2):205–212. doi: 10.1007/s10396-017-0811-8
- Hu Y et al (2019) Automatic tumor segmentation in breast ultrasound images using a dilated fully convolutional network combined with an active contour model. Med Phys 46(1):215–228. doi: https://doi.org/10.1002/mp.13268
- Uppada R, Kodati SP, Rao SK (2021) Automated Computer Aided Diagnosis Using Altered Multi-Phase Level Sets in Application to Categorize the Breast Cancer Biopsy Images. IETE J Res 0(0):1–15. doi: 10.1080/03772063.2021.1962741
- Liu B, Cheng HD, Huang J, Tian J, Liu J, Tang X (2009) Automated Segmentation of Ultrasonic Breast Lesions Using Statistical Texture Classification and Active Contour Based on Probability Distance. Ultrasound Med Biol 35(8):1309–1324. doi: https://doi.org/10.1016/j.ultrasmedbio.2008.12.007
- Huang Q, Miao Z, Zhou S, Chang C, Li X (2021) Dense Prediction and Local Fusion of Superpixels: A Framework for Breast Anatomy Segmentation in Ultrasound Image With Scarce Data. IEEE Trans Instrum Meas 70:1–8. doi: 10.1109/TIM.2021.3088421
- Huang Q, Huang Y, Luo Y, Yuan F, Li X (2020) Segmentation of breast ultrasound image with semantic classification of superpixels. Med Image Anal 61:101657. doi: https://doi.org/10.1016/j.media.2020.101657
- Vupputuri A, Ashwal S, Tsao B, Ghosh N (2020) Ischemic stroke segmentation in multi-sequence MRI by symmetry determined superpixel based hierarchical clustering. Comput Biol Med 116:103536. doi: https://doi.org/10.1016/j.compbiomed.2019.103536
- Ren L, Liu Y, Tong Y, Cao X, Wu Y (2020) Calcification segmentation based on a different scales superpixels saliency detection algorithm. Ultrasound Med Biol 46(12):3404–3412. doi: https://doi.org/10.1016/j.ultrasmedbio.2020.08.013
- Ilesanmi AE, Idowu OP, Makhanov SS (Oct. 2020) Multiscale superpixel method for segmentation of breast ultrasound. Comput Biol Med 125:103879. doi: 10.1016/j.compbiomed.2020.103879
- Hozumi N et al (2005) “Acoustic impedance micro-imaging for biological tissue using a focused acoustic pulse with a frequency range up to 100 MHz,” in IEEE Ultrasonics Symposium, 2005, vol. 1, pp. 170–173, doi: 10.1109/ULTSYM.2005.1602823
- Fadhel MN, Berndl ESL, Strohm EM, Kolios MC (2015) High-Frequency Acoustic Impedance Imaging of Cancer Cells. Ultrasound Med Biol 41(10):2700–2713. doi: https://doi.org/10.1016/j.ultrasmedbio.2015.06.003
- Yap MH et al (Jul. 2018) Automated Breast Ultrasound Lesions Detection Using Convolutional Neural Networks. IEEE J Biomed Heal Informatics 22(4):1218–1226. doi: 10.1109/JBHI.2017.2731873
- Al-Dhabyani W, Gomaa M, Khaled H, Fahmy A (2020) Dataset of breast ultrasound images. Data Br 28:104863. doi: https://doi.org/10.1016/j.dib.2019.104863
- Mittal A, Moorthy AK, Bovik AC (2012) No-Reference Image Quality Assessment in the Spatial Domain. IEEE Trans Image Process 21(12):4695–4708. doi: 10.1109/TIP.2012.2214050
- Mittal A, Soundararajan R, Bovik AC (2013) Making a ‘Completely Blind’ Image Quality Analyzer. IEEE Signal Process Lett 20(3):209–212. doi: 10.1109/LSP.2012.2227726
- Singh A, Kumar S (2020) A novel dice similarity measure for IFSs and its applications in pattern and face recognition. Expert Syst Appl 149:113245. doi: https://doi.org/10.1016/j.eswa.2020.113245
- Bag S, Kumar SK, Tiwari MK (2019) An efficient recommendation generation using relevant Jaccard similarity. Inf Sci (Ny) 483:53–64. doi: https://doi.org/10.1016/j.ins.2019.01.023
- Csurka G, Larlus D, Perronnin F (2013) “What is a good evaluation measure for semantic segmentation?,”
- Moon WK, Shen Y, Bae MS, Huang C, Chen J, Chang R (2013) “Computer-Aided Tumor Detection Based on Multi-Scale Blob Detection Algorithm in Automated Breast Ultrasound Images,” IEEE Trans. Med. Imaging, vol. 32, no. 7, pp. 1191–1200, Jul. doi: 10.1109/TMI.2012.2230403
- Moon WK et al (2020) Computer-aided tumor detection in automated breast ultrasound using a 3-D convolutional neural network. Comput Methods Programs Biomed 190:105360. doi: https://doi.org/10.1016/j.cmpb.2020.105360
Table 1 Quality metric of different techniques together with
|
Enhancement method |
Original |
CLAHE |
AGCWD |
CWS |
SHEHM |
Proposed |
|
BRISQUE |
43.01 |
42.46 |
43.45 |
41.85 |
41.01 |
38.87 |
|
NIQE |
7.744 |
7.38 |
7.69 |
6.53 |
7.57 |
5.57 |
|
Entropy |
7.083 |
6.92 |
6.84 |
6.09 |
6.05 |
4.01 |
Table 2 Similarity metrics between the masked image generated by the techniques against the masked image generated by experts
|
|
DICE |
Jaccard |
BF score |
|
Output image segmented using Superpixel |
0.8595 |
0.7535 |
0.6582 |
|
Output image segmented using AC |
0.8179 |
0.69 |
0.4818 |
|
SHEHM image segmented using AC |
0.7456 |
0.5944 |
0.2624 |
|
CWS image segmented using AC |
0.7866 |
0.6482 |
0.3348 |
|
MSS |
0.8615 |
0.7866 |
0.2792 |
Table 3 Probability of false alarm
|
|
CLAHE |
AGCWD |
CWS |
SHEHM |
Proposed |
|
False alarm Probability |
0.82 |
0.79 |
0.72 |
0.68 |
0.085 |