gms | German Medical Science

Objective: Magnetic resonance imaging (MRI) is the imaging modality of choice for detection of primary cerebral lymphoma. After initiating therapy, follow-up MRI scans are necessary to evaluate treatment response. An automated detection and segmentation of the (residual) lymphoma would provide additional information for the clinician, in particular regarding an increasing workload and physician fatigue. The objective of this study was to evaluate a deep learning model (DLM) based on convolutional neuronal networks for automated detection and segmentation of primary cerebral lymphoma on multiparametric MRI.

Methods: Seventy-nine MRI scans (T1-/T2-weighted, T1-weighted contrast-enhanced (T1CE), T2-weighted FLAIR) from 51 patients with histologically proven primary cerebral lymphomas were selected for this study. Of these scans, six datasets of three patients were excluded due to MR artifacts and three datasets of three patients due to pronounced leukoencephalopathy (Fazekas III). Independent manual segmentations of the tumor core (on T1CE) and peritumoral edema (on FLAIR) by a neurosurgeon and a radiologist served as the ground truth. For this study, a three-dimensional convolutional neural network architecture based on DeepMedic (Biomedia, Singapore) was employed, which was originally established and trained for detection and segmentation of glioblastoma using five-fold cross validation (5-FCV). The DLM did not receive dedicated training for this study. Dice similarity coefficients were calculated to determine segmentation accuracy.

Results: The final data set consisted of 70 MRI scans (45 patients, mean age: 62.5±13.6 years, 22 females). The detection sensitivity of the glioblastoma-trained DLM was 0.89. Compared to the ground truth, the DLM achieved a median dice coefficient of 0.71 for segmentation of tumor core (average size: 11.61 ± 12.87 cm³) and a dice coefficient of 0.74 for peritumoral edema (average size 76.32 ± 63.22 cm³), hence representing a high segmentation accuracy. There were no significant differences of detection and segmentation accuracies between initial and follow-up scans (p>0.05).

Conclusion: Even though primary cerebral lymphoma is a complex and irregular tumor entity, the glioblastoma-trained DLM detects and segments primary cerebral lymphoma with acceptable accuracy on a heterogeneous data set without dedicated training.