Key words CT - semi-automatic - tumor response - interventional oncology - interobserver variability
Introduction
Liver metastases and advanced primary liver tumors, such as hepatocellular carcinomas
(HCCs), are associated with a poor prognosis. Surgical therapeutic options entail
tumor or metastasis resection or liver transplantation (in non-metastasized HCCs).
Local non-surgical therapies, such as transarterial chemoembolization (TACE), selective
internal radiation therapy (SIRT) and radiofrequency ablation (RFA), can be used in
hepatic metastases and HCCs [1 ]
[2 ]
[3 ]
[4 ]. Systemic therapy could constitute classic cytotoxic chemotherapy or targeted treatment
(e. g., sorafenib).
Therapy-induced changes in HCCs and hepatic metastases are classically determined
using the firmly established Response Evaluation Criteria in Solid Tumors (RECIST
1.1) [5 ] or the criteria of the World Health Organization (WHO) [6 ]. These are based on uni- and bidimensional measurements of the entire lesion which
are usually assessed by analyzing transversely oriented images acquired with computed
tomography (CT) or magnetic resonance imaging (MRI). However, the mere quantification
of the size of a lesion has significant limitations, since treatment such as TACE
or SIRT initially results in modified tumor vascularization and not in a size reduction
at the initial stage of tumor response [4 ]
[7 ]
[8 ]. Therefore, the therapeutic effect could be either undetected or underestimated
and lead to inappropriate therapeutic decisions (e. g., unnecessary modification of
the therapeutic regime).
These limitations led to the development of criteria that also account for vascularization
and the extent of possible necrosis. The response criteria according to the European
Association of the Study of the Liver (EASL) are based on bidimensional measurements
of the vital parts of the tumor, i. e., those revealing arterial contrast uptake [9 ]. The RECIST guidelines were also adapted to include and quantify tumor necrosis.
These criteria are known as modified RECIST (mRECIST) and are still based on unidimensional
measurements [10 ]. Both guidelines primarily aimed to assess HCCs but have already been used in the
evaluation of metastases in the context of TACE, SIRT or targeted systemic treatments.
In the clinical routine, radiological assessment of tumor size is usually performed
manually. A major disadvantage of manual measurement is the high intraobserver and
interobserver variability, which may lead to misinterpretation of tumor response [11 ]. Previous studies have shown that the use of semi-automatic measuring techniques
could reduce this variability [12 ], leading to increased precision [13 ] and thus to a more accurate classification of therapeutic response.
The aim of this study was to determine the measurement precision of established and
vitality-based response criteria depending on the method (manual versus semi-automatic)
used to measure hepatic metastases and HCCs under endovascular therapy.
Materials and Methods
Patients
72 consecutive (January 2008 to December 2013) patients (46 male [64 %], 26 female
[36 %]; mean age: 60 years [17 – 83 years]) with HCCs (n = 30 [42 %]) or hepatic metastases
(n = 42 [58 %]) from other tumors (19 colorectal adenocarcinomas, 6 breast adenocarcinomas,
4 malignant melanomas, 3 pancreatic adenocarcinomas, 3 respiratory adenocarcinomas,
3 gastrointestinal neuroendocrine tumors, 4 other tumor types) were included in this
retrospective study.
The inclusion criteria were: (a) transarterial chemoembolization (TACE, n = 29 [40 %])
or (b) selective internal radiation therapy (SIRT, n = 43 [60 %]) of at least one
of the liver lesions. Patients who received LIPIODOL® (Guerbet LLC, Bloomington, Indiana, USA) as part of the TACE regimen were excluded
due to its high attenuation and interference with the evaluation of contrast enhancement.
TACE was conducted by injection of doxorubicin-loaded (50 mg) polyvinyl alcohol particles
(100 µm) into the tumor-feeding segmental liver artery. During SIRT loaded spheres
(30 – 40 µm) with activities between 0.7 and 1.8 GBq were applied.
All patients underwent contrast-enhanced multislice CT (MSCT) before and after local
ablative therapy. Depending on the survival time, the number of procedures (some patients
underwent two or more TACE and/or SIRT) and follow-up CT scans varied. A maximum of
two lesions were included per patient to avoid bias due to a great number of genetically
and presumably phenotypically similar lesions. In the case of more than two lesions,
the two largest ones were chosen.
Written informed consent was obtained from every patient. The study was approved by
the local ethics committee and conducted in accordance with the guidelines of the
institutional review board.
Data acquisition
Images were taken with a 64-slice dual-source CT scanner (SOMATOM® Definition, Siemens AG, Medical Solutions, Forchheim, Germany). The tube voltage
was 120 kV and the collimation was 32 × 0.6 mm. Dose modulation (CARE Dose4 D™ , Siemens AG, Medical Solutions, Forchheim, Germany) was undertaken to reduce radiation
exposure. Iodine-containing contrast agent (Ultravist® -370, Bayer Schering Pharma, Leverkusen, Germany) was injected intravenously at a
constant flow rate of 5 mL/s. The arterial contrast phase was determined dynamically
by means of bolus tracking. The venous contrast phase was defined by a fixed delay
of 75 seconds after i. v. injection. All CT data sets were reconstructed at a slice
thickness of 1.5 mm, using a reconstruction interval of 0.6 mm.
Data preparation
The CT examinations were transferred to a commercially available, dedicated oncological
software suite (mint LesionTM , Mint Medical GmbH, Heidelberg, Germany), working on a server-client principle. The
client, as shown in figure 1, was installed as additional software on dedicated RIS-PACS
workstations. A radiologist (15 years of oncological imaging experience), who was
not involved in further measurements or data processing, identified and tagged the
malignant liver lesions treated with TACE and SIRT on the basis of the images obtained
from the intervention and CT examinations. These liver lesions were identified and
measured independently by two readers (R1 with 4 years and R2 with 1 year of oncological
imaging experience) who were not involved in data preparation. The evaluation was
performed in a semi-random order with a time period of at least 1 month between manual
and semi-automatic evaluation of the same lesion. Both readers were blinded with respect
to the patients’ diseases (i. e., type of liver lesion) and treatments. The abovementioned
software, which supports both manual and semi-automatic radiological measurements,
was used for evaluation purposes.
Manual evaluation
The assessment comprised measurement of the long axis diameter (LAD, [mm]) of the
entire lesion as well as determination of the modified long axis diameter (mLAD, [mm]),
defined as the arterial contrast-enhancing portion and presumably representing the
vital part of the tumor. Both diameters were measured in transversely oriented CT
images, illustrating the largest dimension of the lesion.
Perpendicular to the LAD and mLAD, the shortest diameter of the lesion was determined
and referred to as the short axis diameter (SAD, [mm]) and modified short axis diameter
(mSAD, [mm]), respectively.
Based on these diameters, the areas were calculated by multiplication, resulting in
the WHO (LAD × SAD [mm²]) and the EASL (mLAD × mSAD [mm²]) areas.
Although it is possible to determine the volume manually, this is not feasible in
the daily routine and therefore was not performed.
Semi-automatic evaluation
Semi-automatic two-dimensional (area-based) and three-dimensional (volume-based) segmentation
was performed on all tagged lesions.
The area-based segmentation process was initiated by drawing a circle around the rough
margins of the lesion, preferably in transverse reconstructions. The correct contour
was then approximated based on threshold- and contour-based algorithms ([Fig. 1 ]). Correction tools could be used without restraint to modify any insufficient segmentation
results within a maximum time of 120 seconds per measurement. The area-based LAD,
SAD, WHO, mLAD, mSAD and EASL were computed from these segmentations.
Fig. 1 Example of a semi-automatically determined WHO product with automatically derived
LAD and SAD.Abb. 1 Beispiel einer semiautomatischen Messung nach WHO mit automatisch ermitteltem LAD und
SAD.
For volume-based segmentation, additional contours had to be defined from adjacent
transverse slices or any perpendicular reconstruction planes [14 ]. The volume-based LAD, SAD and WHO were derived from the result. Volume-based EASL
analysis is not supported by the software and therefore was not performed.
All measurement results from every time point were transferred to a dedicated spreadsheet
for further statistical analysis.
Data management
To ensure the comparability of the different parameters, the measurements had to be
converted into standard units as described by James et al. [15 ]. Therefore, all volume and bidimensional measurements were converted into separate
diameters as previously published by different groups [16 ]
[17 ]
[18 ]. These effective diameters were measured in mm and defined as “volume-equivalent
and area-equivalent diameters.” The volume-equivalent diameter (Dvol ) was calculated by inverting the sphere volume formula: Dvol = (6 × V/π)⅓ , whereby V = volume measurement (mm3 ) and Dvol = diameter (mm). To convert bidimensional measurements into unidimensional measurements,
the surface area of a sphere was assumed. By inverting the area formula A = (π × Ds
2 )/4, whereby A = bidimensional measurement (mm2 ) and Ds = diameter (mm), the area-equivalent diameters were calculated using the formula
Ds = 2√((1/π) × A).
Statistical analysis
Statistical analyses were performed using SAS® software, version 9.4, for Windows (SAS Institute, Cary, NC, USA) and IBM SPSS® Statistics 22 for Windows (IBM Corporation, Somers, NY, USA). Inferential statistics
were intended to be exploratory rather than confirmatory. P-values were used to generate
new hypotheses and represent only a metric measure of evidence against the respective
null hypothesis. Thus, neither a global significance level nor local levels were determined,
and no adjustment for multiplicity was made. P-values ≤ 0.05 were considered statistically
significant.
Standard descriptive statistical analyses were performed for the parameters LAD, SAD,
bidimensional WHO, mLAD, mSAD, bidimensional EASL (all manual and semi-automatic)
and volume. Categorical variables are reported as absolute and relative frequencies.
Normally distributed continuous variables are reported as mean ± standard deviation,
and non-normally distributed continuous variables as median (10 %, 90 % quantile).
To assess interobserver variability between readers R1 and R2 for each parameter,
the relative interobserver difference (RID) was determined as RID = | R1 – R2 |/mean
(R1, R2) × 100 %.
To determine absolute agreement between the two readers, intraclass correlation coefficients
(ICC, two-way random single measure) were calculated for manual and semi-automatic
measurements [19 ]. The ICCs ranged from 0 to 1, whereby values from 0.61 to 0.80 indicate substantial
agreement and values from 0.81 to 1 almost perfect agreement.
Results
Lesion characteristics
137 lesions (57 HCCs, 80 metastases) were measured both manually and semi-automatically
on baseline and follow-up CT scans in 72 patients, resulting in a total of 691 observations.
As EASL, mLAD and mSAD are only applicable to lesions with a hypervascularized portion,
fewer lesions were measured in line with these vitality-based criteria.
The medians of the measurements resulting from the manual and semi-automatic methods
differ only slightly from each other, regardless of the number of dimensions taken
into account ([Table 1 ]). Slightly higher deviations can be found in the modified RECIST and EASL criteria.
Table 1
Lesion characteristics and number of measurements for manually and semi-automatically
derived established (long axis diameter [LAD], short axis diameter [SAD], WHO, volume)
and vitality-based (modified LAD [mLAD], modified SAD [mSAD], EASL) parameters. Multidimensional
parameters (WHO, EASL and volume) are given as unidimensional-equivalent diameters
(mm) for better comparability. Measurements are pooled from both readers and all examinations.
Different numbers of measurements result from fewer HCC lesions (established versus
vitality-based parameters) and technical limitations in segmentation (area-based versus
volume-based parameters).Tab. 1 Läsionscharakteristiken und Anzahl der Messungen für manuell und semiautomatisch
bestimmte, etablierte (Längsachsendurchmesser [LAD], Kurzachsendurchmesser [SAD],
WHO, Volumen) und Vitalitäts-assoziierte (modifizierter LAD [mLAD], modifizierter
SAD [mSAD], EASL) Parameter. Mehrdimensionale Parameter (WHO, EASL und Volumen) sind
zwecks Vergleichbarkeit als unidimensionales Äquivalent (mm) angegeben. Messungen
über beide Reader und alle Untersuchungen. Unterschiedliche Anzahl an Messungen durch
geringere Anzahl von HCCs (etablierte versus Vitalitäts-assoziierte Parameter) und
technische Limitationen der Segmentierung (flächenbasiert versus volumenbasierte Parameter).
Lesion size in mm (mm-equivalent for WHO, EASL and volume)
Median (10 %, 90 % quantile), n = number of measurements
Parameter
Manual
Semi-automatic Area-based
Semi-automatic Volume-based
LAD (mm)
33.8 (15.0, 90.0), n = 691
35.3 (16.4, 90.1), n = 691
34.8 (16.4, 86.5), n = 669
SAD (mm)
26.6 (12.3, 69.2), n = 690
27.4 (12.7, 70.3), n = 691
25.9 (12.2, 65.2), n = 669
WHO (mm)
33.8 (15.2, 88.4), n = 690
35.0 (16.2, 89.3), n = 691
33.9 (15.5, 84.2), n = 669
mLAD (mm)
27.2 (12.0, 59.7), n = 206
29.4 (13.9, 61.1), n = 212
mSAD (mm)
18.6 (7.7, 40.3), n = 206
21.3 (8.6, 43.1), n = 212
EASL (mm)
25.1 (10.7, 51.7), n = 206
27.7 (12.5, 55.7), n = 212
Volume (mm)
27.8 (12.8, 64.6), n = 626
Relative interobserver difference (RID)
The RID ([Table 2, ]
[Fig. 2 ]) – as a measure of divergence between readers R1 and R2 – reveals no statistically
significant difference in the established parameters (LAD, SAD, WHO), regardless of
the measurement technique, i. e., manual LAD 6.0 % and semi-automatic area-based LAD 5.9 %,
manual SAD 7.7 % and semi-automatic area-based SAD 6.9 %.
Table 2
Relative interobserver difference (RID) between reader 1 and reader 2 for each manual
and semi-automatic area-based parameter (long axis diameter [LAD], short axis diameter
[SAD], WHO, volume, modified LAD [mLAD], modified SAD [mSAD], EASL). The RID in the
established parameters (LAD, SAD, WHO) reveals no statistically noticeable difference,
whereas the RID for vitality-based parameters is lower in the case of semi-automatic
measurement.Tab. 2 Relative Interobserverdifferenz (RID) zwischen Reader 1 und 2 für jeden manuell und
semiautomatisch bestimmten flächengeleiteten Parameter (Längsachsendurchmesser [LAD],
Kurzachsendurchmesser [SAD], WHO, Volumen, modifizierter LAD [mLAD], modifizierter
SAD [mSAD], EASL). Die RID der semiautomatisch bestimmten, etablierten Parameter (LAD,
SAD, WHO) unterscheidet sich von den manuellen Messungen nicht statistisch signifikant,
wohingegen die RID der Vitalitäts-assoziierten Parameter (mLAD, mSAD, EASL) deutlich
niedriger ist als die der jeweilig manuellen Messungen.
Relative interobserver difference in % (reader 1 vs. reader 2)
Median (10 %, 90 % quantile)
Parameter
Manual
Semi-automatic Area-based
LAD
6.0 (1.1, 21.4)
5.9 (1.0, 20.8)
SAD
7.7 (0.9, 24.7)
6.9 (1.0, 27.8)
WHO
5.7 (0.8, 19.1)
5.4 (0.8, 20.2)
mLAD
12.5 (2.1, 31.8)
3.4 (0.4, 8.5)
mSAD
12.7 (1.9, 52.5)
5.7 (0.7, 15.6)
EASL
10.4 (1.4, 45.9)
1.8 (0.4, 6.4)
Volume
4.1 (0.5, 17.0)
Fig. 2 Box plots of the relative interobserver difference (RID) between reader 1 and reader
2 for each manual and semi-automatic area-based and volume-based parameter (long axis
diameter [LAD], short axis diameter [SAD], WHO, volume, modified LAD [mLAD], modified
SAD [mSAD], EASL). A larger RID and more outliers are found in the vitality-based
parameters (mLAD, mSAD, EASL) when determined manually; this can be counteracted by
using a semi-automatic area-based approach.Abb. 2 Boxplots der relativen Interobserverdifferenz (RID) zwischen Reader 1 und 2 für jeden
manuell und semiautomatisch bestimmten flächen- und volumenabgeleiteten Parameter
(Längsachsendurchmesser [LAD], Kurzachsendurchmesser [SAD], WHO, Volumen, modifizierter
LAD [mLAD], modifizierter SAD [mSAD], EASL). Eine größere RID und mehr Ausreißer können
bei den manuell bestimmten Vitalitäts-assoziierten Parametern (mLAD, mSAD, EASL) gefunden
werden; dies kann durch ein semiautomatisches Messverfahren reduziert werden.
In contrast, the deviation in the vitality-based criteria (mLAD, mSAD and EASL) is
lower in the semi-automatic area-based measurements compared to manual measurements,
i. e., manual mLAD 12.5 % and semi-automatic area-based mLAD 3.4 %, manual EASL 10.4 %
and semi-automatic area-based EASL 1.8 %. Moreover, the number of outliers is drastically
reduced using the semi-automated area-based method of measurement ([Fig. 1 ]).
The volume can only be determined semi-automatically and has no manually derived equivalent.
Its median deviation of 4.1 % is relatively low compared to the other parameters.
Intraclass correlation coefficient (ICC)
The ICC – as an indicator of interobserver agreement – is consistently high for the
established parameters LAD, SAD and WHO, with no relevant difference between manual
and semi-automatic area-based measurements ([Table 3 ]).
Table 3
Intraclass correlation coefficients (ICC) (two-way random single measure) and 95 %
confidence intervals for agreement between reader 1 and 2 in manual and semi-automatic
measurements (long axis diameter [LAD], short axis diameter [SAD], WHO, volume, modified
LAD [mLAD], modified SAD [mSAD], EASL). Semi-automatic, area-based determination of
vitality-based parameters (mLAD, mSAD, EASL) leads to substantially higher agreement
(ICC) between reader 1 and 2 compared to manual measurements of the same parameters.Tab. 3 Intraclass Korrelationskoeffizient (ICC) (two-way random single measure) und 95 %
Konfidenzintervalle der Übereinstimmung zwischen Reader 1 und 2 bezüglich manueller
und semiautomatischer Messungen (Längsachsendurchmesser [LAD], Kurzachsendurchmesser
[SAD], WHO, Volumen, modifizierter LAD [mLAD], modifizierter SAD [mSAD], EASL). Semiautomatische,
flächenabgeleitete Bestimmung Vitalitäts-assoziierter Parameter (mLAD, mSAD, EASL)
führt zu einer substanziell höheren Übereinstimmung (ICC) zwischen Reader 1 und 2
im Vergleich zu einer manuellen Bestimmung der selben Parameter.
Intraclass correlation coefficient (reader 1 vs. reader 2)
ICC (95 % CI), n = number of measurements
Parameter
Manual
Semi-automatic Area-based
Semi-automatic Volume-based
LAD
0.984 (0.980, 0.987), n = 324
0.982 (0.976, 0.986), n = 324
0.976 (0.969, 0.982), n = 303
SAD
0.975 (0.969, 0.984), n = 323
0.958 (0.948, 0.966), n = 324
0.758 (0.706, 0.802), n = 303
WHO
0.984 (0.980, 0.987), n = 323
0.978 (0.973, 0.983), n = 324
0.903 (0.878, 0.923), n = 303
mLAD
0.897 (0.846, 0.932), n = 85
0.997 (0.996, 0.998), n = 105
mSAD
0.844 (0.770, 0.896), n = 85
0.992 (0.988, 0.995), n = 105
EASL
0.875 (0.815, 0.917), n = 85
0.998 (0.997, 0.999), n = 105
Volume
0.987 (0.984, 0.990), n = 284
The ICCs from manual measurement of mLAD, mSAD and EASL are lower. The manual parameter
with the best correlation, mLAD, has an ICC of 0.897, for example.
Taking the type of lesion – HCCs versus metastases – into account ([Table 4 ]), there are only small differences regarding the LAD (all ICCs above 0.95). A lower
ICC can be found in the SAD of HCCs, especially in the semi-automatic 3 D measurements
(3 D SAD 0.780).
Table 4
Intraclass correlation coefficients (ICC) by tumor entity (two-way random single measure)
and 95 % confidence limits for agreement between reader 1 and 2 in manual and semi-automatic
measurements (long axis diameter [LAD], short axis diameter [SAD], WHO, volume, modified
LAD [mLAD], modified SAD [mSAD], EASL). The mLAD, mSAD and bidimensional EASL were
not applicable to metastases and have therefore been omitted. The ICC is equally high
for LAD regardless of the measurement method.Tab. 4 Intraclass Korrelationskoeffizient (ICC) nach Tumorentität (two-way random single
measure) und 95 % Konfidenzintervalle der Übereinstimmung zwischen Reader 1 und 2
bezüglich manueller und semiautomatischer Messungen (Längsachsendurchmesser [LAD],
Kurzachsendurchmesser [SAD], WHO, Volumen, modifizierter LAD [mLAD], modifizierter
SAD [mSAD], EASL). Die Parameter mLAD, mSAD, und bidimensionaler EASL sind per Definition
nicht auf Metastasen anwendbar und entfallen daher in der Betrachtung. Der ICC für
LAD ist über alle Messverfahren konstant hoch.
Intraclass correlation coefficient (reader 1 vs. reader 2)
ICC (95 % CI)
HCC
Metastases
Parameter
Manual (n = 131)
Semi-automatic Area-based (n = 132)
Semi-automatic Volume-based (n = 129)
Manual (n = 192)
Semi-automatic Area-based (n = 192)
Semi-automatic Volume-based (n = 174)
LAD
0.962 (0.946, 0.973)
0.957 (0.938, 0.970)
0.962 (0.945, 0.974)
0.992 (0.989, 0.994)
0.991 (0.988, 0.993)
0.988 (0.983, 0.991)
SAD
0.856 (0.937, 0.969)
0.898 (0.860, 0.927)
0.780 (0.693, 0.843)
0.984 (0.979, 0.988)
0.989 (0.985, 0.992)
0.747 (0.673, 0.806)
WHO
0.965 (0.950, 0.975)
0.943 (0.921, 0.960)
0.817 (0.741, 0.871)
0.992 (0.990, 0.994)
0.994 (0.992, 0.995)
0.971 (0.961, 0.978)
Discussion
Manual radiological measurements in CT examinations are an established clinical approach
and form the basis for any evaluation of imaging in oncology. Nevertheless, numerous
recent studies have demonstrated lower interobserver variability and higher reproducibility
with semi-automatic measurements [12 ]
[13 ]
[17 ]
[20 ]
[21 ]
[22 ]
[23 ]
[24 ]. These advantages permit more reliable and accurate classification of the therapeutic
response and directly influence treatment decisions. These studies are limited in
that they mostly focused on the relatively easy task of lung nodule [17 ]
[22 ]
[23 ] or lymph node segmentation in CT examinations [12 ]
[13 ].
The segmentation of liver lesions in MRI examinations is also firmly established and
usually involves a semi-automatic, volume-based approach [25 ]
[26 ]. On the other hand, reliable segmentation of liver lesions in CT – with its lower
soft-tissue-contrast – is a more demanding task and has been addressed only recently
[24 ]
[27 ]. Special challenges are posed by the initially variable morphology which changes
over the course of new targeted und endovascular therapies due to decreased tumor
vascularization with subsequently reduced contrast enhancement or even necrosis. In
light of these unavoidable hindrances, the mode of measurement (manual or semi-automatic)
should not add any further uncertainty.
Our data reveal a consistently high level of measurement precision (reflected by the
ICC) for any semi-automatically derived measurements, including the vitality-based
parameters mLAD, mSAD and EASL. In contrast, the precision of the manual measurements
of these vitality-based parameters is considerably lower. As the ICC does not mainly
depend on the number of cases, this could be explained at least in part by the smaller
area to be measured with a consecutively higher variation.
One possible explanation for the higher ICCs of the semi-automatically derived measurements
is that the standardized semi-automatic workflow offers guidance (e. g. by proposing
reconstruction planes or boundaries) in difficult situations, counteracting the lesion-
and therapy-dependent variations and leading to less variation.
This advantage is not expected to come to the fore in the relatively easy task of
generally determining lesion size. Our data ultimately reveal no relevant differences
in precision between manual and semi-automatic measurements for the established parameters
LAD, SAD and WHO, regardless of the lesion type (HCCs versus metastases).
In this regard our results are consistent with previous studies that report a higher
ICC for semi-automatic CT measurements of lymph nodes [13 ] and pulmonary nodules [22 ]
[23 ] and extend the applicability to the semi-automatic evaluation of liver lesions in
CT. Analogous results were published recently with a focus on MRI evaluation of liver
lesions after intra-arterial therapy [28 ]
[29 ].
For the probably more demanding CT segmentation and measurement, a current publication
evaluated HCCs under systemic molecular-targeted therapy [30 ]. In contrast to this study, we applied therapy (TACE or SIRT) selectively to the
liver arteries that could make a difference in the homogeneity and intensity of the
therapy effects, making measurements even more difficult.
As an additional benefit, a semi-automatic workflow facilitates standardized and complete
documentation [31 ]. This helps reduce measurement time in follow-up examinations by a third, compensating
for the slightly longer, initial segmentation time [32 ]. Furthermore, it offers a systematic overview and guidance in patients with multiple
examinations, possibly at different sites, thereby permitting monitoring of a diversified
therapeutic spectrum.
Limitations
This study is limited to the extent that the measured liver lesions were not excised.
Thus, the actual size and, depending on this, the accuracy could not be determined.
However, a surgical intervention would not have been justified, and the precision
– as a relative measure – is not influenced by our approach, which is accepted for
in-vivo studies [17 ].
Furthermore, we chose a single-center, retrospective study design. Because the focus
of interest was the measurement agreement between the readers, each measurement was
regarded as independent, thus potentially disregarding correlations between lesions
in the same patient and between different time points.
We did not evaluate the mean segmentation time which could pose a bias due to over-accurate
editing of contours. To prevent this we restricted the maximum segmentation time to
120 seconds per lesion [24 ].
The CT scanner and reconstruction protocols were kept constant to the disadvantage
of limited generalizability.
Conclusion
We conclude that vitality-based tumor measurements of hepatocellular carcinomas and
metastases after transarterial local therapies should be performed semi-automatically
due to greater measurement precision, thus increasing the reproducibility and, in
turn, the reliability of therapeutic decisions. Manual and semi-automatic measurements
of established parameters offer the same level of precision, but preference should
be given to the semi-automatic approach due to the possibility of generating systematic
documentation.
