H Ziyaee¹*, C Cardenas², D Yeboa¹, j Li¹, J Johnson¹, Z Zhou¹, J Sanders¹, R Mumme¹, L Court¹, T Briere¹, J Yang¹, (1) UT MD Anderson Cancer Center, Houston, TX, (2) The University of Alabama at Birmingham, Birmingham, AL

• 
  H Ziyaee

SU-H430-IePD-F5-2 (Sunday, 7/10/2022) 4:30 PM - 5:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: To develop an automated framework for contouring brain metastases on MRIs using a deep learning model to aid treatment planning for stereotactic radiosurgery (SRS), and to thoroughly understand its limitations.

Methods: An nnU-Net model was trained to auto-contour brain metastases on post-contrast 3D T1-weighted MR images from SRS patients. The MRI dataset included an average number of 4.0±3.5 metastases per patient, with average metastasis size of 10.0±7.4mm. The dataset was randomly split into 845 patients for training and 206 patients for testing (80:20). Metrics including positive predictive value (PPV), sensitivity, and dice similarity coefficient (DSC) were used to quantitatively assess the model. False-positives (FPs) were investigated by reviewing radiology notes and follow-up imaging. Two CNS radiation oncologists scored 24 patients using a 5-point scale on clinical acceptability.

Results: The PPV, sensitivity, and DSC were 92.1±16.4%, 88.6±17.8%, and 82.2±9.5% respectively on patient-averaged. Considering only large metastases (≥6mm), these metrics were 96.1±17.1%, 94.5±18.1%, and 86.8±7.5%, respectively on patient-averaged which shows near perfect PPV and sensitivity with high segmentation performance. An investigation into FPs indicated that 9% were true-positive (TP) metastases that were not identified on the original radiology review but were identified on subsequent follow-up imaging. 54% were real metastases (TP) that were identified but not contoured mainly because patient underwent whole brain radiotherapy pre/post SRS treatment. Only 37% were real FPs metastases. Clinical reviews showed that 75% of the auto-contour TPs were clinically acceptable and could be used without any modification. The remaining 25% needed minor editing only.

Conclusion: We developed a robust automated framework to detect metastases for brain MRIs. This model achieved a high detection performance with a high level of clinical acceptability. It can provide consistent and accurate contouring for clinical application with an early detection ability.

Funding Support, Disclosures, and Conflict of Interest: Brockman Foundation

Keywords

Segmentation, MRI, Brain

Taxonomy

IM/TH- image Segmentation: MRI

Contact Email