Automated follicle tracking facilitates standardization and may improve work flow

Ultrasound is used to identify pathology and to confirm normality. This requires a subjective interpretation of the image display, which can be modified according to the object or area of interest. Objective assessment of an ultrasound image requires some form of measurement to be made, which should be performable in a reproducible manner and provide a valid result. Ultrasound, therefore, is open to interpretation and is dependent upon the observer regardless of whether a subjective or an objective examination is being made. Automatic data analysis has the potential to remove any observer bias and to reduce the time needed for measurements, but it must be both valid and reliable. This is particularly true within the field of reproductive medicine, with ultrasound being used on a daily basis to follow the ovarian response to gonadotropins in a process known as ‘follicle tracking’1. Each follicle is identified and displayed in its maximum diameter. Two-dimensional measurements are then made and their mean is taken as the true follicular diameter. This is relatively straightforward but follicle tracking becomes more difficult as the number of follicles increases. A typical patient undergoing controlled ovarian stimulation as part of assisted reproduction treatment would be expected to develop somewhere between five and ten follicles within each ovary, although it is not unusual for there to be many more. The validity and reliability of measurements of follicular diameter are likely to reduce as the number of follicles increases, because it becomes increasingly difficult to ensure that every follicle is accounted for. The time required for these measurements also increases as the number of follicles increases, which may adversely affect measurement reliability. Automated measurement of follicular size could potentially address both of these issues and introduce a standard for such measurements2-4. SonoAVC (Automatic Volume Calculation: GE Medical Systems, Zipf, Austria) is a new software program that identifies and quantifies hypoechoic regions within a three-dimensional dataset and provides automatic estimation of their absolute dimensions, mean diameter and volume. Each individual volume is given a specific color and the automated measurements of its mean diameter (relaxed sphere diameter), its maximum dimensions (x, y, z diameters) and its volume are displayed by these colors in descending order of size. An unlimited number of volumes can theoretically be quantified, which makes it an ideal tool for assessment of the ovaries in women undergoing controlled ovarian stimulation. Automated volume measurements require the acquisition of a three-dimensional dataset. This is then displayed using the multiplanar view and a region of interest is selected by manually moving a box to fit the maximum proportions of the ovary. Once the dataset has been correctly positioned, SonoAVC is implemented. The automated analysis takes around 6 s and the individual follicles are then displayed together with their dimensions and relative sizes (Figure 1). Our preliminary experience has shown that the software has great potential and that it provides both reliable and valid measurements of follicles. In a study comparing the automated measurements to those derived from measuring the follicular aspirate5, we found that SonoAVC estimated the volume of a follicle to within ± 0.5 cm3. This was significantly more accurate than were volume measurements derived using algorithms (the sphere and ellipsoid formulae) based on the mean follicular diameter calculated by both two- and three-dimensional techniques. Follicular diameter measurements were also more reliable when made with SonoAVC, which provides two different mean diameters6. The ‘relaxed sphere diameter’ is derived from the automatic volume calculation by assuming that the follicle is spherical and applying the formula used for the estimation of the volume of a sphere. The ‘x, y, z mean diameter’ is calculated by taking the mean of the three perpendicular diameters. The software is formulated to first measure the maximum follicular diameter in the longitudinal or transverse plane and then to measure the two diameters perpendicular to this. This technique was based on preliminary information from a UK questionnaire looking at the role of ultrasound in fertility treatment, which specifically asked users, including ultrasonographers, nurse specialists and doctors, to describe diagrammatically how they would measure a follicle (unpubl. data). Ultimately, measurements of diameter are relatively arbitrary as no non-uniform three-dimensional object actually has a true diameter. Volume is therefore a more appropriate measure. However, two-dimensional measurements are used in clinical practice as they are easy to perform and three-dimensional data acquisition is a relatively new technique that is not yet widely used for this purpose. Three-dimensional ultrasound imaging using SonoAVC software, within a stimulated ovary during in-vitro fertilization treatment. The color-coded follicles and their individual measurements are visible. The validity of the automated volume measurements is not due to the improved reliability of follicular diameter measurements. The algorithm used to calculate volume is actually based on the summation of all the volume elements within any given hypoechoeic region. A volume element, or ‘voxel’, is a three-dimensional pixel and therefore provides true volumetric information even if part of this is based on computation from the reconstructed image planes within an acquired dataset. The volume given by SonoAVC is calculated from the total number of voxels within the ultrasound image and can be converted into standard units because each voxel has a given size. SonoAVC can be used to assess the whole ovary and is most useful towards the latter stages of controlled ovarian stimulation, when there are multiple follicles of varying sizes. Preliminary studies have shown that it counts the number of follicles and provides estimates of their dimensions much more reliably than does an experienced observer using real-time two-dimensional ultrasound. Furthermore, the time taken for such analysis is significantly less with SonoAVC, as one would expect. Preliminary findings in a study of 30 subjects undergoing assessment on the day that human chorionic gonadotropin was administered to promote oocyte maturation prior to oocyte retrieval, showed that the mean time for measurement of both ovaries per subject was 180.5 ( ± 63.6) s for SonoAVC and 236.1 ( ± 57.1) s for real-time two-dimensional assessment (P < 0.01). The time taken for the automated measurements included the time for data acquisition 39.0 ( ± 6.0) s and data analysis 141.5 ( ± 64.1) s. These data show that three-dimensional ultrasound can significantly reduce the time of the ultrasound examination, which has important implications for work flow. Theoretically, technicians could be used to analyze data if automated analysis can be shown to be both valid and reliable, and to perform measures as well as or better than conventional manual measures. Clearly, this needs to be considered in the context of the clinical scenario, but, if acceptable and following a well-defined training program, it could be introduced into clinical practice and monitored by cyclical audit. There are limitations to SonoAVC, however, and it is important to recognize them. Some follicles are not identified following the initial automatic assessment, some follicles are identified incorrectly and in most cases there is some extraovarian artifact. There are several options for post-processing that require manual or automatic modification of the initial output. Manual techniques may be used for single follicles but the automated options are applied to the dataset as a whole. The most simplistic tool involves clicking on individual follicles that have not been identified as well as those that have been included incorrectly. These areas then become colored and uncolored, respectively, and only colored areas have their measurements displayed. The next step involves tracing around follicles to identify them as separate or single entities, although this may also be achieved by using the ‘separation’ and ‘growth’ functions that modify how each follicle is color-filled. The latter option is good when there are aberrances across the whole ovary, but in most cases editing is required for a few follicles and the manual technique is preferable. Our initial work has shown that the software only performs well if the original dataset is of adequate quality. Image quality affects two key areas when the whole ovary is considered: the ovarian cortex and follicular margins. If the ovarian cortex is not clearly defined, the software tends to consider bowel gas and extraovarian fluid as a follicle. This can be remedied easily by using the edit features, which allow the user to manually define follicles and to confirm whether they are single or multiple follicles. The user can trace around a ‘true follicle’ and this is then identified as a separate entity. The extraovarian tissue incorrectly recognized as a follicle can then be removed by clicking on it. This takes time and post-processing is required in the majority of cases. The time required for post-processing depends on image quality, which can be assessed qualitatively by considering how much of the ovarian contour can be seen (Table 1). We have used a similar technique for the application of inversion mode, which is also dependent on image quality, and both techniques perform best when more than 90% of the contour can be seen (Grade 1) but work adequately if more than 50% can be identified (Grade 2)7. The software works less well when less than 50% of the ovarian contour is readily evident (Grade 3). Ultimately, this is a limitation of all ultrasound techniques and also affects measurements using two-dimensional ultrasound. One way to avoid post-processing is to use Virtual Organ Computer-aided AnaLysis (VOCAL™, GE Medical Systems) to define the ovarian contour8. If SonoAVC is applied after the cortex has been defined, the inappropriate inclusion of extraovarian tissue is greatly reduced. Indistinct follicular margins are much harder to address. Generally, if the cortex is well defined, the dataset is likely to be of good quality and the follicles will be seen as separate structures. The main problem with poorer quality images is the coalescence of follicles that are truly individual. Post-processing is not always able to address this and a number of datasets cannot be assessed accurately. Manual measurements are required in these cases and, whilst this adds time to the overall assessment, many of the follicles are correctly identified and measured automatically. Ultrasound has continued to evolve since its inception and image quality is likely to continue improving, particularly with the integration of computer-based systems and software development. SonoAVC can be used to quantify and track ovarian cysts and to count the number of antral follicles within a normal (Figure 2) or a polycystic (Figure 3) ovary, although the software performs less well with smaller follicles, for which there is a need for more extensive post-processing. Obstetric applications include assessment of early pregnancy through calculation of the gestational and yolk sac volumes (Figure 4) and demonstration of the fetal brain cavities and fluid-filled abdominal organs during the later stages of pregnancy. Automatic volume calculation of antral follicles in an unstimulated normal ovary. Follicles measuring 2–10 mm are displayed. Automatic volume calculation of antral follicles in an unstimulated polycystic ovary. Follicles measuring 2–10 mm are displayed. Three-dimensional ultrasound imaging using automatic volume calculation software, showing a singleton fetus at 10 weeks' gestation (C-plane, lower left image) within its gestational sac (colored purple and measuring 20.84 mL) and separate from the yolk sac (colored green and measuring 0.02 mL). SonoAVC was developed by GE Medical Systems in association with K plus Competence Center (Advanced Computer Vision) and part funded by the K plus Program.