The 109 pulmonary nodules were classified by a consensus of two radiologists according their size, relation with surrounding lung parenchyma and neighbouring structures (nodule type), structure (density) and margins based on visual analysis of the 1.25mm thick axial images.
1) size (maximum axial diameter) : 0,9% <5mm (n=1); 28,4% >5< 10 mm (n= 31); 35,7% >=10<15 mm (n= 39); 12,8% >015<18 mm (n=14); 22% >18<30 mm (n=24) (Table 2).
2) relation with surrounding lung parenchyma and neighbouring structure (also called type): well circumscribed (n= 16); iuxta-vascular (n= 67); iuxta-pleural (n=13); mixed (n.13) iuxta-pleural and iuxta-vascular (Table 3).
3) density of the nodules, 70/109 pulmonary nodules were solid, 8 were non solid (ground glass opacity) and 31 nodules were partly solid (13 of them had solid part + ground glass, 18 of them with solid part + excavation) (Table 4).
4) margins: smooth (indicated as 1, n=55), spiculated (indicated as 2, n= 22) and mixed (partly spiculated, indicated as 3, n= 11) (Table 5).
Volume determination was possible 865 times out of 965 reconstruction (89,63%). Segmentation of the nodules was not possible in 100/965 reconstructions (10,3%) and in 12/93 nodules detected after intravenous administration of contrast medium it was not possible to obtain volumetric data because of failure in segmentation process.
We evaluated the different causes of the failure of segmentation, and consequently of volumetric measurements:
strict relation between nodule and vascular structures
strict relation between nodule and pleura
partial volume artefact: it happened in nodules with max diameter of 4-18 mm and reconstructed at slice thickness setting of 2,5 mm and 5 mm, Full FOV and FOV 10.
impossibility of the software to identify lesions after intravenous administration of contrast medium (9/93 nodules)
interruption of the segmentation process because of the over-plus of data.
Statistical results
Table 6 shows the RMSE% obtained from MANOVA analysis of error contrasts between each technique and the average one. In the most of techniques an RMSE% for volume lower than maximum diameter was obtained. The most accurate techniques were 125F10 (RMSE%=7.8%) for maximum diameter and 125S (RMSE%=6.5%) for volume. Technique MDC supplied poor accuracy both for volume (RMSE%=21.7%) and maximum diameter (RMSE%=15.4%).
It can be noted as four techniques, 125B, 125S, 25B and 125F10, allowed the measurability of 100-101 nodules, out of 109. Considering very important nodule measurability and observing that the accuracy of these N (N=4) techniques was sufficiently high for both parameters (in the worst case we had an RMSE% equal to 11.2%), we selected them for calculating the mean reference value, mN , and the corresponding relative errors (see eqn. 1), εk, of any k technique.
Table 7 reports the correction coefficient, αk (k = 1, 2, …10), for nodule measurements, and the coefficients, β(lv)k and β(uv)k, for the definition of lower and upper values of 95% CI for both parameters and all techniques. We can observe that the most precise (no corrections) and accurate technique, with the highest number of measurable nodules, is the 125S. In fact: both its α coefficients are equal to 1, meaning that neither diameter maximum nor volume measurements must be corrected; it exhibits narrow 95% confidence intervals indicating good accuracy: they respectively range from 88% to 112% and from 86% to 114% of measured value. Although slight corrections have to be made, similar performances are supplied by technique 125B with even better accuracy for volume estimation.
Table 7 allows the use of any of the ten techniques for the estimation of the true nodule dimension. For example, suppose we want to estimate the volume of a pulmonary nodule, measured with technique 5B to be 1000 mm3. The correction factor corresponding to technique 5B for volume measurements is α5B(volume) = 0.91, while the coefficients associated to 95% CI are β(lv)5B(volume)=0.46 and β(lv)5B(volume)= 1.36; therefore the true volume value is estimated of 910 mm3 with a 95% CI ranging from 460 to 1360 mm3. The wide 95% CI indicate poor accuracy of technique 5B.
For what concerns the MANOVA between-nodules analysis, multivariate statistical differences (Wilks’ lambda test) were found only for the interactions between measurement techniques and margin factor. No differences were found for nodule density and type. Therefore, monovariate ANOVA analysis of relative errors was performed for investigating differences in margin characteristics.
Table 8 shows the measurement techniques giving statistical significant results. It can be noted that technique 125S, among the best four techniques, gives a little but statistically significant overestimation (4.8%) of maximum diameter of nodules having margin type 3. Consequently, when maximum diameter of nodules with margin type 3 is measured by this technique, it could appear convenient its systematic reduction of 4.8%. However, a not reported investigation about this correction showed a marked error variance of these nodules which supplied only slight improvements in the measurement accuracy: the 95% CI got narrower of only just 2%. Considering also the too little sample size of nodules with margin equal to 3, which does not guarantee robust statistical results, we therefore omitted this correction.
As expected, technique 5B confirms to be problematic, giving high margin-dependent underestimates (minus sign) of maximum diameter, large volume overestimates (plus sign) in margin 1 (14.9%) and 2 (12.2%) and underestimates in margin 3 (-11.5%). In this case, as in the others techniques showing higher margin dependent errors, suitable corrections for margin type would surely give greater advantages. Nevertheless, we opted to avoid margin correction because the sample size was too little for applying suitable cross-validation procedures to test correction validity.
In support of our choice to neglect margin-dependent corrections, Table 7 clearly shows as margin-independent global correction anyway allows a decrease of Table 8 errors, especially for technique 5B strongly biased to maximum diameter underestimation and volume overestimation.
Short considerations
Relationship between algorithm of reconstruction, characterization and volumetric measurement of the lung nodule.
There is not a statistically significant difference between Bone and Standard retro-reconstruction algorithm in volumetric measurement of pulmonary nodule with LVCAR software, at least using slice thickness 1,25 mm. (Fig. 8)
Relationship between slice thickness, characterization and volumetric measurements of pulmonary nodule
The use of lower or higher slice thickness (respect to 1.25), using the same acquisition parameters (FOV, reconstruction interval) seems to be less accurate and to give worst results; in fact it gives a overestimation of the measurements at slice thickness > 1.25 and a underestimation at slice<1.25.
Moreover, it misinterprets density of the nodule; for examples, using Full FOV, at 5 mm slice thickness, iuxtavascular solid nodules (max. diameter <10 mm) are automatically recognized by the software like non solid or partly solid and the outcome volume is bigger than the real ones. It may happens because of some artifacts (partial volume artifatc); in fact, with an high slice thickness the shape of a small pulmonary nodule is undefined and irregular and the LVCAR software interprets all the nodules or the peripheral portion of it like a non-solid lesion (Fig. 9).
Relationship between FOV and volumetric measurements of the nodule
There is not a statistically significant difference between FullFOV and FOV10mm in volumetric measurement of pulmonary nodule with LVCAR software, at least using slice thickness 1,25 mm [reference 19].
Effect of intravenous contrast medium administration and volumetric measurements of the nodule
Technique with MDC supplied poor accuracy both for volume (RMSE%=21.7%) and maximum diameter (RMSE%=15.4%) (Table 5) (Fig. 10) [reference 20].