Automated Detection and Segmentation of Brain Metastases in Malignant Melanoma: Evaluation of a Dedicated Deep Learning Model

BACKGROUND AND PURPOSE

Malignant melanoma is an aggressive skin cancer in which brain metastases are common. Our aim was to establish and evaluate a deep learning model for fully automated detection and segmentation of brain metastases in patients with malignant melanoma using clinical routine MR imaging.

MATERIALS AND METHODS

Sixty-nine patients with melanoma with a total of 135 brain metastases at initial diagnosis and available multiparametric MR imaging datasets (T1-/T2-weighted, T1-weighted gadolinium contrast-enhanced, FLAIR) were included. A previously established deep learning model architecture (3D convolutional neural network; DeepMedic) simultaneously operating on the aforementioned MR images was trained on a cohort of 55 patients with 103 metastases using 5-fold cross-validation. The efficacy of the deep learning model was evaluated using an independent test set consisting of 14 patients with 32 metastases. Manual segmentations of metastases in a voxelwise manner (T1-weighted gadolinium contrast-enhanced imaging) performed by 2 radiologists in consensus served as the ground truth.

RESULTS

After training, the deep learning model detected 28 of 32 brain metastases (mean volume, 1.0 [SD, 2.4] cm³) in the test cohort correctly (sensitivity of 88%), while false-positive findings of 0.71 per scan were observed. Compared with the ground truth, automated segmentations achieved a median Dice similarity coefficient of 0.75.

CONCLUSIONS

Deep learning-based automated detection and segmentation of brain metastases in malignant melanoma yields high detection and segmentation accuracy with false-positive findings of <1 per scan.

Differentiation of Hemorrhage from Iodine Using Spectral Detector CT: A Phantom Study

BACKGROUND AND PURPOSE

Conventional CT often cannot distinguish hemorrhage from iodine extravasation following reperfusion therapy for acute ischemic stroke. We investigated the potential of spectral detector CT in differentiating these lesions.

MATERIALS AND METHODS

Centrifuged blood with increasing hematocrit (5%-85%) was used to model hemorrhage. Pure blood, blood-iodine mixtures (75/25, 50/50, and 25/75 ratios), and iodine solutions (0-14 mg I/mL) were scanned in a phantom with attenuation ranging from 12 to 75 HU on conventional imaging. Conventional and virtual noncontrast attenuation was compared and investigated for correlation with calculation of relative virtual noncontrast attenuation. Values for all investigated categories were compared using the Mann-Whitney U test. Sensitivity and specificity of virtual noncontrast, relative virtual noncontrast, conventional CT attenuation, and iodine quantification for hemorrhage detection were determined with receiver operating characteristic analysis.

RESULTS

Conventional image attenuation was not significantly different among all samples containing blood (P > .05), while virtual noncontrast attenuation showed a significant decrease with a decreasing blood component (P < .01) in all blood-iodine mixtures. Relative virtual noncontrast values were significantly different among all investigated categories (P < .01), with correct hemorrhagic component size estimation for all categories within a 95% confidence interval. Areas under the curve for hemorrhage detection were 0.97, 0.87, 0.29, and 0.16 for virtual noncontrast, relative virtual noncontrast, conventional CT attenuation, and iodine quantification, respectively. A ≥10-HU virtual noncontrast, ≥20-HU virtual noncontrast, ≥40% relative virtual noncontrast, and combined ≥10-HU virtual noncontrast and ≥40% relative virtual noncontrast attenuation threshold had a sensitivity/specificity for detecting hemorrhage of 100%/23%, 89%/95%, 100%/82%, and 100%/100%, respectively.

CONCLUSIONS

Spectral detector CT can accurately differentiate blood from iodinated contrast in a phantom setting.