Abstract
Left ventricular non-compaction (LVNC) is defined by an increase of trabeculations in left ventricular endo-myocardium. Although LVNC can be in isolation, an increase in hypertrabeculation often accompanies genetic cardiomyopthies. Several enhancements are proposed and implemented to improve a software tool for the automatic quantification of the exact hyper-trabeculation degree in the left ventricular myocardium for a population of patients with LVNC cardiomyopathy (QLVTHC-NC). The software tool is developed and evaluated for a population of 18 patients (133 cardiac images). An end-diastolic cardiac magnetic resonance images of the patients are the input of the software, whereas the left ventricular mass, volumes and proportion of trabeculation produced by the compacted zone and the trabeculated zone are the outputs. Significant improvements are obtained with respect to the manual process, so saving valuable diagnosis time. Comparing the method proposed with the fractal proposal to differentiate LVNC and non-LVNC patients in subjects with previously diagnosed LVNC cardiomyophaty, QLVTHC-NC presents higher diagnostic accuracy and lower complexity and cost than the fractal criterio.
Introduction
Cardiomyopathies are a heterogeneous group of genetic diseases that impair cardiac muscle structure and function and can lead to heart failure. Cardiomyopathies are incurable and relatively frequent and constitute a major cause of morbidity and mortality. Left ventricular non-compaction (LVNC) has been defined by the The American Heart Association as a distinct primary genetic cardiomyopathy caused by an arrest of the normal compaction process of the developing myocardium1, whereas the European Society of Cardiology refers to LVNC as an unclassified cardiomy-opathy2 because LVNC could appear as a morphological trait of other cardiomyopathies, especially hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM). Whether LVNC is a distinct cardiomyopathy or a mor-phologic trait shared by different types of cardiomyopathies is still debated3,4,5. Besides, acquired and even reversible forms of hypertrabeculation can also occur and have been described mainly in endurance athletes and pregnant women.
LVNC patients show prominent trabeculations with deep recesses within the myocardium6. LVNC is caused by defects in the process of compaction by which trabeculae integrate in the ventricular wall, giving rise to the smooth inner surface of the postnatal ventricle7.
There is controversy about the accuracy of different methods for quantifying the non-compacted myocardium7 and the timing of measurement (end-systole or end-diastole)5. One of the main problems of non-compaction is the lack of clear diagnostic criteria that differentiate those patients who will really develop a cardiomyopathy from those who will not present events at follow-up. The development of a diagnostic algorithm or score is required to establish the diagnostic criteria.
The estimation of the non-compacted myocardium by endocardium delineation from cardiac magnetic resonance (CMR) images has been proposed to quantify the non-compacted mass. The traditional technique involves the manual delineation of the different structures by an expert cardiologist, requiring a significant amount of time and being dependent on the observer’s point of view8.
Recently, Captur et al.9,10 proposed a fractal analysis about endocardial borders to produce a continuous variable, the fractal dimension (FD), which measures their complexity. The result of the analysis is expressed as the value of the slope of the integral of the succession of measurements of pixels in a grid and the compacted myocardium is not evaluated. The integration in hospitals of the fractal analysis is very difficult due to the high cost of the software license and FD does not represent a completely reliable measurement. Besides, FD has not been correlated with clinical outcomes.
In previous works11,12 we proposed and evaluated the first automatic software tool based on medical experience that accurately quantifies the degree of left ventricle (LV) hyper-trabeculation (QLVT) for a dilated cardiomyopathy series, using the cardiac images obtained by magnetic resonance (MR). In a recent work13 we presented an updated software tool for the automatic quantification of the exact hyper-trabeculation degree in the left ventricle myocardium for a population of Hypertrophic Cardiomyopathy (QLVTHC) patients. End-diastolic CMR images of the coronal slices of the heart of the patients are the input of the software, while the volumes of the compacted zones and the trabeculated zones are necessary to produce the percentage quantification of the trabecular zone with respect to the compacted zone. The tool is capable of adapting to different situations in populations of HCM and DCM patients.
Our main goal is to obtain the percentage quantification of trabeculation in the LV myocardium in a serie of patients with previously diagnosed LVNC cardiomyophaty (QLVTHC-NC). Manual tuning by cardiologists of different parameters is required to identify the LV cavity, detect the trabecular zones in the interior of the LV cavity and establish accurately the external layer of the myocardium. Therefore, some improvements with regard QLVTHC are proposed, such as the detection of different MSERs in a centered ROI to determine automatically the LV cavity, the accurate and reliable detection of the RV cavity at any location, the refination of the search to obtain the external layer of the myocardium and the trabecular zones and the processing of different slices in a reverse order allowing the proposed method to process patients with different cardiomyopathies efficiently. The software gives understandable clinical measurements of areas, volumes and masses of the compacted and non-compacted left ventricular myocardium through the application of machine vision algorithms. This software tool is based on medical experience and it has been tested by different experts in cardiac image analysis. Therefore, this computationally assisted method can save valuable diagnosis time compared with the traditional processing, so minimizing the possibility of human appreciation error. Moreover, the population of patients with LVNC cardiomyopathy will be processed by the fractal analysis9,10 in order to compare to QLVTHC-NC to determine the better diagnostic accuracy.
The rest of the paper is organized as follows. Section 2 outlines the improvements of the computing method proposed (QLVTHC-NC). Several experiments to test the proposal are presented in section 3. A comparison between QLVTHC-NC and the method proposed by Captur9,10 is performed in section 4. Finally, section 5 summarizes the work, concludes the paper and introduces future work.
Methods
Based on the QLVT algorithm described in11, we developed an optimized algorithm13, presented in Algorithm 1, for the automatic quantification of the degree of LV hyper-trabeculation in Hypertrophic Cardiomyopathy patients (QLVTHC). Several enhancements complemented a tool to process dilated and hypertrophic patients efficiently.
Algorithm 1.
Optimized algorithm for the automatic quantification of the LV hyper-trabeculation degree in DCM and HCM patients (QLVTHC)
| 1: for each slice of a patient do |
| 2: Detect different MSERs in a centered ROI, identify LV cavity and apply Convex Hull |
| 3: Identify RV cavity |
| 4: Detect external layer of compacted zone |
| 5: Detect trabecular zones |
| 6: Compute areas of trabecular zones and the compacted zone in the LV |
| 7: Obtain percentage quantification of the trabecular zone with respect to the compacted area |
| 8: end for |
| 9: Compute volumes and masses of trabecular zones and the compacted zone. Obtain percentage quantification |
LVNC cardiomyopathy could be revealed in many patients. Therefore, the LV cavity, the RV cavity and the external layer of the myocardium can present irregular features or different structures to DCM or HCM. Thus, different enhancements are proposed:
The integration of DICOM format to read the input images, which are the different slices obtained for a particular patient based on end-diastolic CMR images (without any annotation by cardiologists). The thickness of each slice (in mm), the spacing between slices (in mm) and the pixel spacing is automatically determined
The different MSERs are detected in a centered ROI (the size of the ROI is the 60% of the completed image size) of each input image by the use of OpenCV14. As the LV cavity is normally represented by a circular shape, the centroid of each MSER detected is computed in order to identify automatically the LV cavity anywhere in the image and for applying Convex Hull15.
A second optimized search process to obtain the external layer and the trabecular zones, thanks to the previous application of Convex Hull, is performed. Now in the implementation different lines are drawn between the centroid of the LV cavity and the possible space where the points of the external layer are found. The lengths of different lines to search the external layer depend on a parameter called e-expand, which enables us to reach the external layer of the myocardium. We have optimized this parameter taking into account the properties of LVNC cardiomyopathy.
The accurate and reliable detection of the RV cavity at any location of a slice.
The possibility to appear the slices of a patient in reverse order (from basal to apical), obtaining the total volumes of the trabecular zone and the compacted zone in a automatic way.
These enhancements complete a tool to process LVNC cardiomyopathies efficiently.
The output of the algorithm allows cardiologists to quantify the exact trabeculation degree in the LV quickly and automatically for a determined patient with LVNC cardiomyophaty (QLVTHC-NC).
Results
We applied the computing method presented in section 2 to a population of 18 patients (identified from L1 to L18) with previously diagnosed definite LVNC cardiomyopathy, confirmed by either genetic testing or family aggregation. Each patient contains different cardiac images or slices obtained by magnetic resonance. Magnetic resonance studies were performed with a 1.5 T scanner (Avanto: Siemens, Germany) in the Hospital Vall d’Hebron (HVH) in Barcelona (Spain). The images were obtained in synchronisation with the ECG and in apnea. The left ventricle function was evaluated with balanced steady-state free precession (b-SSFP) sequences (repetition interval of 3.8 ms., echo time of 1.7 ms., flip angle of 60◦, matrix of 224 × 224, echo train length of 23, slice thickness of 8 mm, slice gap of 2 mm, with 20 phases).
All participants were recruited from an inherited Cardiomyopathy Clinic. Patients with an available good quality CMR study were included from a referral centre. The first CMR test was used for the study. Clinical and outcome data prospectively collected in a database was available for analysis. Left ventricle trabecular myocardial mass (TM) and compacted mass were measured (compacted myocadium, CM) using the software proposed in the before section (QLVTHC-NC) for automatic delineation of borders. Mean CM was 77.3 ± 25.8 g and TM 38.6 ± 14.1 g. Mean percentage of trabeculated myocardium (TM%) from total was 33.1% ± 4.6%.
In Figure 1, the volumes of compacted zone (VCZ) and trabecular zone (VT) for 18 patients (133 slices or cardiac images) with previously LVNC diagnosed cardiomyopathy are presented. There are great differences among the values of the volumes, showing QLVTHC-NC is capable of detecting volumes of several sizes. Once the volumes are obtained, the percentage quantification of trabeculated myocardium (TM%), presented in Figure 2, is easily computed and the masses are calculated using the established density.
Figure 1:
Volumes of compacted zone (VCZ) and trabecular zone (VT) computed by QLVTHC-NC for patients (L1-L18)
Figure 2:
Percentage quantification of trabeculated myocardium (TM%) computed by QLVTHC-NC for patients (L1-L18)
0.1. Quality evaluation by cardiologists
To assess the performance of the QLVTHC-NC, an expert evaluation on the output images with the identified zones was performed by skilled cardiologists. These experts graded the images, using the scale proposed by Gibson et al.16 (Table 1).
Table 1:
Evaluation scale to measure the diagnostic quality of medical images
| 5.0 Exact match: there is no noticeable differences |
| 4.5 |
| 4.0 Noticeable differences: they are not diagnostically significant |
| 3.5 |
| 3.0 Small diagnostically significant differences |
| 2.5 |
| 2.0 Significant diagnostic information is lost |
| 1.5 |
| 1.0 Large diagnostically significant differences |
For each slice of a patient, the cardiologists observed the original image and the output of the QLVTHC-NC method. Table 2 shows the mean score of cardiologists to the identification of the different heart parts (LV cavity, trabecular zones and compacted zone) performed by the proposed method.
Table 2:
Results of mean quality measure obtained from cardiologists for images with LV cavity, trabecular zones and compacted zone identified by QLVTH-NC
| 5.0 | 4.5 | 4.0 | 3.5 | 3.0 | 2.5 | 2.0 | 1.5 | 1.0 | |
| QLVTHC-NC | 87.97% | 7.52% | 4.51% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
The different improvements introduced in QLVTHC-NC reveal the visibility is very clear in all cases. For all the sequences experimented, the cardiologists were able to detect the presence of diagnostic features clearly, and an exact measure of the volumes and masses of compacted zone and trabecular zone can be obtained in a short time. 100% of the evaluated images have no noticeable differences for diagnosis (scores ≥ 4.0). Moreover, most of the images tested are exact (87.97%) and do not vary by even one pixel according to the cardiologists (scores = 5.0). Moreover, the rest of the slice tested (12.03%) are practically exact for a correct diagnosis (scores = 4.5 or = 4.0). Therefore, QLVTHC-NC is ready to operate and adapt to different degree of LVNC cardiomyopathy. Moreover, the tool help to issue several diagnoses and to analyse the different heart parts in diverse positions and the source of cardiac images come from an external hospital.
In Figures 3 and 4 the output slices of the QLVTHC-NC are presented for L5 and L6, respectively. In each output slice, the compacted zone is delimited between the external layer (in grey) and the trabeculated zone (in blue), whereas the LV cavity without trabecules is also delimited (in red). L5 and L6 have been previously diagnosed with LVNC cardiomyopathy and the slices appear in a reverse order, LV is not always in the center of the slice and the pixels-value are very different between them.
Figure 3:
(a)-(h) Slices 1 to 9 of L5
Figure 4:
(a)-(h) Slices 1 to 8 of L6
Cardiologists can observe the trabeculated and the compacted zone in a simple and precise way, obtaining an accurate degree of the percentage quantification of trabeculated myocardium in order to establish the appropriate diagnosis All these slices of these patients have been evaluated by cardiologists with a 5.0, so demonstrating QLVTHC-NC is able to adapt and detect different forms of compacted zones and trabecular zones accurately and positions in a slice, a reverse order for the slices to compute the volumes of compacted and trabecular zones, distinct degree of LVNC cardiomyopathy (TM% of L5 is 47.35 and L6 is 33.98) and very different pixels-value.
Therefore, QLVTHC-NC provides easy and automatic delineation of borders to obtain clinical measurements of volumes and masses of the compacted and non-compacted LV myocardium in diverse situations, positions and distinct degree of LVNC cardiomyopathy, by adjusting the different parameters. The final output is obtained in less than 10 seconds per slice, thus the processing of a patient can be performed usually in less than 1 or 2 minutes. This time is insignificant compared with the manual process traditionally used, where cardiologists need around 20 minutes per slice to delimit, measure and compute the quotient of the thickness between the trabecular area and the compacted zone for all segments.
Comparison
We have executed the method developed by Captur9,10 to the population of 18 patients (identified as L1 to L18) with previously diagnosed LVNC cardiomyophaty, obtained by magnetic resonance studies performed in the hospital mentioned in section 3. We have executed both methods on a 4-core Intel® CPU i7-4700MQ with Hyperthreading running at 2.40GHz.
In both proposals the execution time required to process each slice of the input can be less than 1 second. Therefore, a patient can be processed in real time. Figure 5 shows both the percentage quantification of trabeculated myocardium
Figure 5:
Results of TM% and global FD10 for patients L1 to L18
(TM%) obtained by QLVTHC-NC and the global FD obtained by Captur method for 18 patients.
The optimum diagnostic threshold for global FD LV was 1,2610, whereas the cut-off for the percentage quantification of trabeculated myocardium (TM%) was 27.4%12 to differentiate LVNC and non-LVNC patients. Therefore, all patients have been marked correctly as LVNC by QLVTHC-NC. However, of the 18 patients, four (L4, L6, L11 and L18) has been mislabelled as health or normal by Captur method. This is due to the compacted myocardium is not evaluated in this method and the global FD represents the tortuosity measured by pixilation of the line of the endocardial borders. In the top of the Figure 6 the areas of compacted zone (ACZ) and trabecular zone (AT) are showed for the patients L5 and L6, respectively, whereas in the bottom of Figure 6, the percentage quantification of trabeculated myocardium (TM%) for each slice of L5 and L6 are presented. As we can observe in the top of the Figure 6, the sizes of the trabeculated and compacted zone may vary depending on a given slice and is considered by QLVTHC-NC to determine the percentage quantification of trabeculated myocardium for each slice and patient. However, the Captur proposal only takes into account the tortuosity of the left ventricle cavity to determine the fractal dimension, establishing a misdiagnosis in 22.22% of the patients analysed.
Figure 6:
Area of compacted zone (ACZ) and trabecular zone (AT) for each slice of the patients L5 (a) and L6 (b). Percentage quantification of trabeculated myocardium (TM%) for each slice of the patients L5 (c) and L6 (d)
Table 3 shows both the FD and the TM% of these patients for each slice divided in basal (B1, B2, B3), mid (M1, M2, M3) and apical (A1, A2, A3) thirds. The Captur proposal also computed the maximal FD in the basal, mid and apical thirds and the optimum diagnostic threshold for maximal FD was 1,3010 to differentiate LVNC and non-LVNC thirds of slices. Now, the identification is not correct as LVNC slice by the Captur method in the M1, M2, A1, A2 and A3 slices of L4, in B1, B2, B3, M1, A1, A2 and A3 slices of L6, in B1, B2 and B3 slices of L11 and in A1, A2, A3 and B3 slices of L18. This local diagnostic marker for LVNC solves and determines as LVNC the mid third of L4. However, the apical third of L4, the basal, mid and apical thirds of L6, the basal third of L11 and the apical and basal thirds of L18 continue marked as non-LVNC, obtaining an accuracy lower than QLVTH-NC.
Table 3:
Results of FD10 and TM% (QLVTHC-NC) for each slice of patients L4, L6, L11 and L18
| Basal | Mid | Apical | |||||||
| Patient | B1 | B2 | B3 | M1 | M2 | M3 | A1 | A2 | A3 |
| L4-FD | 1,13 | 1,21 | 1,35 | 1,27 | 1,28 | 1,31 | 1,24 | 1,06 | 1,11 |
| L4-TM% | 0,81 | 24,77 | 33,29 | 44,75 | 46,19 | 49,17 | 51,89 | 43,59 | 30,38 |
| L6-FD | 1,22 | 1,26 | 1,23 | 1,18 | 1,18 | 1,21 | 1,17 | 1,27 | |
| L6-TM% | 48,14 | 43,79 | 37,44 | 35,61 | 25,12 | 31,16 | 45,73 | 30,33 | |
| L11-FD | 1,17 | 1,23 | 1,20 | 1,33 | 1,33 | 1,21 | 1,20 | 1,22 | |
| L11-TM% | 35,44 | 36,20 | 31,02 | 41,35 | 34,06 | 25,64 | 18,21 | 18,87 | |
| L18-FD | 1,19 | 1,17 | 1,17 | 1,18 | 1,25 | 1,14 | |||
| L18-TM% | 33,83 | 36,73 | 34,61 | 26,01 | 22,46 | 35,93 | |||
On the other hand, both tools labelled as non-LVNC slices B1 and B2 of L4, M3 of L6, A1, A2 and A3 of L11 and A1 and A2 of L18 demonstrating that health slices can be differentiated of LNVC cardiomyopathy.
Moreover, the high cost of the license of the fractal method per year by the complexity of this tool has prevented reproducibility by other medical centers or hospitals. Conversely, QLTVHC-NC could be used and integrated in medical centers on request.
Conclusions
We have proposed several improvements for a tool to quantify the hypertrabeculation in the left ventricle of the myocardium. The proposal is based on the automatic delineation of the left ventricular endocardial and epicardial borders from CMR diastolic images to perform and adapt efficiently for a population of 18 patients with LVNC cardiomyopathy, reducing the diagnosis time considerably and minimizing the possibility of human appreciation error. The input tested are the end-diastolic CMR images of the coronal slices of the heart of the patients (133 slices), which come from an external hospital.
Understandable clinically relevant values of trabeculated and compacted myocardium (areas, volumes and masses) are produced by adjusting different parameters. Several outputs are showed with LVNC cardiomyopathy to demonstrate the efficiency of the tool. The percentage quantification of the trabeculated myocardium varied from 28.62 to 47.35, demonstrating that QLVTHC-NC is able to detect both insignificant and very important trabecular zones and compacted zones accurately and consistently in patients with different degree of LVNC cardiomyopathy. We have demonstrated that the tool can be used and integrated in an external hospital (HVH).
In subjects with previously diagnosed cardiomyophaties, the percentage quantification of the trabeculated myocardium measured by QLVTHC-NC distinguishes LVNC from health with higher diagnostic accuracy than the biological signal obtained by fractal analysis. Moreover, QLVTHC-NC presents lower complexity and cost than the method developed by Captur.
The tool can be extended to the exact quantification of the trabecular zones in the right ventricle and therefore, could be used and integrated in hospitals in order to accelerate diagnoses by cardiologists.
Acknowledgements
This work was supported by the Spanish MCIU and AEI, as well as European Commission FEDER funds, under grant RTI2018-098156-B-C53.
Figures & Table
References
- 1.Maron BJ, Towbin JA, Thiene G, Antzelevitch C, Corrado D, Arnett D, et al. Contemporary Definitions and Classification of the Cardiomyopathies. Circulation. 2006;113(14):1807–1816. doi: 10.1161/CIRCULATIONAHA.106.174287. [DOI] [PubMed] [Google Scholar]
- 2.Rapezzi C, Arbustini E, Caforio A, Charron P, Gimeno-Blanes J, Heliö T, et al. Diagnostic work-up in cardiomy- opathies: bridging the gap between clinical phenotypes and final diagnosis. A position statement from the ESC Working Group on Myocardial and Pericardial Diseases. European Heart Journal. 2013;34(19):1448–1458. doi: 10.1093/eurheartj/ehs397. [DOI] [PubMed] [Google Scholar]
- 3.Brouwer WP, Germans T, Head MC, van der Velden J, Heymans MW, Christiaans I, et al. Multiple myocardial crypts on modified long-axis view are a specific finding in pre-hypertrophic HCM mutation carriers. European Heart Journal - Cardiovascular Imaging. 2012;13(4):292–297. doi: 10.1093/ehjci/jes005. [DOI] [PubMed] [Google Scholar]
- 4.Olivotto I, d’Amati G, Basso C, Van Rossum A, Patten M, Emdin M, et al. Defining phenotypes and disease pro- gression in sarcomeric cardiomyopathies: contemporary role of clinical investigations. Cardiovascular Research. 2015;105(4):409–423. doi: 10.1093/cvr/cvv024. [DOI] [PubMed] [Google Scholar]
- 5.Paterick TE, Umland MM, Jan MF, Ammar KA, Kramer C, Khandheria BK, et al. Left Ventricular Noncom- paction: A 25-Year Odyssey. Journal of the American Society of Echocardiography. 2012;25(4):363–375. doi: 10.1016/j.echo.2011.12.023. [DOI] [PubMed] [Google Scholar]
- 6.Jenni R, Oechslin EN, van der Loo B. Isolated ventricular non-compaction of the myocardium in adults. Heart. 2007;93(1):11–15. doi: 10.1136/hrt.2005.082271. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Towbin JA, Lorts A, Jefferies JL. Left ventricular non-compaction cardiomyopathy. Lancet. 2015;386(9995):813–825. doi: 10.1016/S0140-6736(14)61282-4. [DOI] [PubMed] [Google Scholar]
- 8.Jacquier A, Thuny F, Jop B, Giorgi R, Cohen F, Gaubert JY, et al. Measurement of trabeculated left ventricular mass using cardiac magnetic resonance imaging in the diagnosis of left ventricular non-compaction. European Heart Journal. 2010;31:1098–1104. doi: 10.1093/eurheartj/ehp595. [DOI] [PubMed] [Google Scholar]
- 9.Captur G, Muthurangu V, Cook C, Flett AS, Wilson R, Barison A, et al. Quantification of left ventricular trabec- ulae using fractal analysis. Journal of Cardiovascular Magnetic Resonance. 2013 May;15(1):36. doi: 10.1186/1532-429X-15-36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Captur G, Lopes LR, Patel V, Li C, Bassett P, Syrris P, et al. Abnormal Cardiac Formation in Hypertrophic Cardiomyopathy-Fractal Analysis of Trabeculae and Preclinical Gene Expression. Circulation: Genomic and Precision Medicine. 2014;7(3):241–248. doi: 10.1161/CIRCGENETICS.113.000362. [DOI] [PubMed] [Google Scholar]
- 11.Bernabé G, Cuenca J, de Teruel PEL, Giménez D, Gonza´lez-Carrillo J. A software tool for the automatic quantification of the left ventricle myocardium hyper-trabeculation degree. Procedia Computer Science. 2015 June;51:610–619. [Google Scholar]
- 12.Bernabé G, González-Carrillo J, Cuenca J, Rodŕıguez D, Gimeno-Blanes JR. Performance of a New Software Tool for Automatic Quantification of Left Ventricular Trabeculations. Revista Espan˜ola de Cardiolog´ıa (English Edition) 2017;70(5):405–407. doi: 10.1016/j.rec.2016.07.006. [DOI] [PubMed] [Google Scholar]
- 13.Bernabé G, Casanova JD, Cuenca J, Gonza´lez-Carrillo J. A self-optimized software tool for quantifying the degree of left ventricle hyper-trabeculation. Journal of Supercomputing. 2019;75(3):1625–1640. [Google Scholar]
- 14.OpenCV: Open Source Computer Vision Library 2017. https://opencv.org/
- 15.Sklansky J. Finding the Convex Hull of a Simple Polygon. Pattern Recogn Lett. 1982 Dec;1(2):79–83. [Google Scholar]
- 16.Gibson D, Spann M, Woolley SI. A wavelet-based region of interest encoder for the compression of angiogram video sequences. Information Technology in Biomedicine, IEEE Transactions on. 2004 June;8(2):103–113. doi: 10.1109/titb.2004.826722. [DOI] [PubMed] [Google Scholar]