PURPOSE

Quantitative measurements of lesion volume and distribution have significant value for clinicians evaluating disease progression. Subjective measurements based on clinician’s visual inspection and manual segmentation are often vulnerable to inter- and intra-rater variability, resulting in low reproducibility. Thus, objective automated lesion segmentation tools have been developed to overcome these problems. Since the appearance of lesions varies across MRI protocols, incorporating different MRI scans provides more information for lesion segmentation tools when delineating lesions. In current clinical practice, a 3D T1-weighted scan acquired with 1.2 mm slices is recommended for volumetric analysis of global measures and change of important structures over time. T2 FLAIR imaging is recommended for evaluating brain lesions since T2 FLAIR is the most sensitive measure of white matter lesion in the cerebrum. Therefore, combining the T1-weighted and T2 FLAIR data could result in the optimal lesion segmentation. LesionQuant, a NeuroQuant product (CorTechs Labs Inc., San Diego, California, USA), is a fully-automated lesion segmentation tool for clinical use. It was designed to provide accurate and reproducible lesion segmentations using both T1-weighted and T2 FLAIR data. In this study, objective methods were implemented to evaluate LesionQuant’s lesion segmentation performance.

METHODS

All T1 and T2 FLAIR scans have 1.2 mm slice thickness and in-plane pixel size of 1.0 mm x 1.0 mm. Figure 1 shows a schematic overview of the LesionQuant lesion segmentation. The details of the processing are described as follows: 1) T1-weighted and T2 FLAIR images were loaded into LesionQuant. Pre-processing corrections and intensity normalization were performed. 2) T1 and T2 FLAIR data were aligned and normalized. 3) Lesion segmentation was performed. The thresholds for lesion segmentation were set based on T1 and T2 FLAIR data. 4) Regional lesion classification was performed using segmentation information provided by T1 data. 5) Output segmentation results. In order to avoid the system bias, selected participants had total lesion volumes greater than 5 cc, and a lesion distribution including at least 5 lesions outside the periventricular region.Accuracy test: 3D T2 FLAIR and 3D T1-weighted scans of 31 subjects were acquired with GE 3T scanners and processed through LesionQuant. All 31 brains were manually segmented by an expert. Accuracy evaluation was performed on a voxel to voxel basis after LesionQuant outputs and expert manual segmentation were spatially registered. The Pearson’s coefficient, Intraclass Correlation Coefficient (ICC), Dice coefficient were used as objective evaluation metrics. Reproducibility test: LesionQuant reproducibility was defined as how well LesionQuant handled different MRI brain scans for the same patients. The assumption was made that minimal brain anatomy changes occurred within one month for the cross-scanner analysis, and were essentially identical for the same session analysis. Repeated 3D T1 and 2D/3D FLAIR images were acquired using four different scanners (Toshiba 3T Titian, Toshiba 1.5T Titan, GE 1.5T Signa and Philips 1.5T Achieva). All scans had 1.0 mm x 1.0 mm in plane pixel size and 1.2 mm slice thickness. Agreement of the total lesion volume between the first and the second scan for each subject was evaluated utilizing Dice coefficient and absolute volume difference.

RESULTS

For the total lesion segmentation, the accuracy test showed Pearson’s correlation coefficient was 0.98, ICC was 0.94 and Dice coefficient was 83.4% ± 7.5%. Reproducibility analysis showed that Dice coefficient was 89.0% ± 3.5% and absolute difference was 1.0% ± 0.8%.

DISCUSSION

Initial results demonstrated good overall reproducibility for repeat scans. The accuracy results showed high agreement between LesionQuant and expert manual segmentation as shown in Figure 2 and Figure 3. There are several limitations to the analyses conducted in this initial study. The number of participants included in this study is relatively small when considering the variability of lesion presentation in MS and other white matter diseases. In addition, further studies should focus on the inclusion of additional MS participants in order to capture the full spectrum of lesion distributions and disease severity. Future analyses should also incorporate patients with a variety of diseases that can present with T2 FLAIR lesions. Future analyses will also compare LesionQuant software with automatic and semi-automatic research tools that are freely available for lesion segmentation.

CONCLUSION

The initial results of this study demonstrate that the accuracy of LesionQuant is comparable to expert manual segmentation of lesions and had high reproducibility. This suggests that LesionQuant could be a useful tool in quantitative analysis of the brain lesions.

Acknowledgements

No acknowledgement found.