Automatic segmentation model of primary central nervous system lymphoma based on multiple sequences of magnetic resonance images using deep learning

Abstract

Abstract Purpose and Background. Accurate quantitative assessment of PCNSL by gadolinum-contrast Magnetic resonance imaging (MRI) is closely related to therapy planning, surveillance and prognosis, However, precise volume assessment by manual segmentation is time-consuming and subject to high intra- and interrater variabilities by imaging readers, with poor consistency. In order to solve this problem, we constructed a multimodal artificial intelligence deep learning segmentation model based on multi-sequence MRI images of patients with PCNSL and identified its efficiency, so as to quantitatively calculate volume and other indicators, and compare the consistency and accuracy with doctors' labeling methods. Materials and Methods. A total of 41 PCNSL patients from six Chinese medical centers with pathologically confirmed PCNSL were analyzed. Region of interest (ROI) was manually segmented on contrast-enhanced T1-weighted and T2 scans. Fully automated voxelwise segmentation of tumor components was performed using a 3D convolutional neural network (DeepMedic) trained on gliomas (n = 220). deep-learning model (DLM) segmentations were compared to manual segmentations performed in a 3D voxelwise manner by two readers (radiologist and neurosurgeon; consensus reading) from T1 CE and FLAIR, which served as the reference standard. The Dice similarity coefficient (DSC) were used to evaluate the performance of the models. Successful detection of PCNSL was defined if the DLM obtained a spatial overlap with the manual segmentation of the tumor core (at least one voxel, DSC >0). Mann-Whitney U test was applied to compare continuous variables, while chi-squared test was used for categorical variables between groups. A two-tailed P value <0.05 indicated statistical significance. Results. The DLM detected 66 of 69 PCNSL, representing a sensitivity of 95.7%. Compared to the reference standard, DLM achieved good spatial overlap for total tumor volume (TTV, union of tumor volume in T1 CE and FLAIR; average size 77.16 ± 62.4 cm3, median DSC: 0.76) and tumor core (contrast enhancing tumor in T1 CE; average size: 11.67 ± 13.88 cm3, median DSC: 0.73). High volumetric correlation between automated and manual segmentations was observed (TTV: r= 0.88, P < 0.0001; core: r = 0.86, P < 0.0001). Performance of automated segmentations was comparable between pre-treatment and follow-up scans without significant differences (TTV: P = 0.242, core: P = 0.177). Conclusion. Compared to manual segmentation on routine clinical MRI images, our automatic segmentation model of PCNSL based on multiple sequences of MRI images displayed comparable segmentation in both pretherapy and the process of the treatment, despite the complex and multifaceted appearance of this lymphoma subtype , implying its immense potential to be used in the whole follow-up monitoring process of PCNSL.