MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Detection and Segmentation of Brain Metastases On MR Images Using Machine Learning and a Novel Optimized Thresholding Technique

D Hsu*, A Ballangrud, L Cervino, J Deasy, A Li, H Veeraraghavan, M Hunt, A Shamseddine, M Aristophanous. Memorial Sloan-Kettering Cancer Center, New York, NY

Presentations

(Sunday, 7/12/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: To automatically detect and segment brain metastases on MR images for radiotherapy, while providing the physician with an estimate of segmentation confidence.


Methods: Over 400 patients treated with stereotactic radiosurgery (SRS) were retrospectively identified for training and testing data. All patients were imaged before receiving radiation with a spoiled-gradient 3D T1 MR sequence post-Gadolinium injection. The gross tumor volume (GTV) contours, authored by physicians for treatment planning, were taken as the reference contours. There were 3.1 ± 1.4 lesions per patient, with diameters between 0.2 and 3 cm. The MR images and the reference contours were used to train a V-Net 3D convolutional neural network and to obtain 3D probability maps. A novel aspect of the process is that two independent confidence thresholds were used for detection (potential tumor voxels) and segmentation (final tumor voxels) to attain a high detection efficiency and the most accurate GTV segmentations. The Dice similarity coefficient (DSC) and Hausdorff distances were used to evaluate the quality of the segmentations.


Results: The algorithm’s performance was estimated using five-fold cross-validation. At a 70% confidence threshold, the algorithm detects brain metastases with an efficiency of 81 ± 4 %, and a false positive rate of 1.7 ± 0.7 lesions per patient. The patient Dice coefficient was 0.70 ± 0.11. For a successfully detected lesion, the predicted GTV had a Dice coefficient of 0.78 ± 0.07 and a Hausdorff distance of 2.7 ± 1.2 mm compared to the physician’s contour. Adjusting the detection threshold allows for either higher detection efficiency or fewer false positives.


Conclusion: This work provides a flexible and accurate algorithm which analyzes a single MRI sequence to identify and segment brain metastases. A confidence estimate for each predicted GTV is thought to help inform the physician in clinical practice.

Keywords

Segmentation, Brain, MRI

Taxonomy

IM/TH- image Segmentation: MRI

Contact Email