gms | German Medical Science

Objective: Cerebral metastases represent an advanced stage of malignant melanoma and requires an adjustment of the oncological treatment concept. Magnetic resonance imaging (MRI) is the imaging modality of choice for detection and follow-up of brain metastasis. Owing to an increasing workload, physician fatigue with the inherent risk of misdiagnosis is a relevant concern, which can affect patient diagnosis, treatment and outcome. The purpose of this study was to develop and evaluate a deep learning model (DLM) based on convolutional neuronal networks that detects and segments brain metastases fully automatically on multiparametric MRI from diverse referring institutions, scanners and vendors in patients with malignant melanoma.

Methods: A total of 54 multiparametric MRI scans (T1-weighted, T1-weighted contrast-enhanced, T2-weighted, Fluid-attenuated inversion recovery sequences) from 54 patients with brain metastasis (mean age: 63.54 ± 13.83 years, 24 females) were used to train and validate the DLM. Independent manual segmentations of the metastases was performed in a voxel-wise manner by two experienced radiologists and served as the ground truth. A three-dimensional convolutional neural network architecture based on DeepMedic (Biomedia, Singapore), which was originally established for detection and segmentation of glioblastoma, was used and underwent dedicated training using five-fold cross validation (5-FCV).

Results: At the time of initial diagnosis, the patients had 102 brain metastases. The average metastasis volume was 2354.27 ± 7809.15 mm³, as determined by the ground truth. Before 5-FCV, the glioblastoma-trained DLM achieved a detection rate of 0.47 (median dice coefficient: 0.64). After 5-FCV, the sensitivity of the DLM increased to 0.87 (p<0.05), with a corresponding median dice coefficient of 0.74 (p<0.05).

Conclusion: After dedicated training using 5-FCV, our DLM detects brain metastases of malignant melanoma in MRI with high accuracy and achieves sufficient segmentation rates despite diverse scanner data. Therefore, it may pose a valuable adjunct for radiologists in cancer imaging.