Computer Aided Detection Tool for Difficult-to-Detect Brain Metastases

Unmet Need

An estimated 20% of all cancer patients will develop brain metastases in their lifetime. These metastases are first diagnosed by neuroradiologists using CT, and more commonly, MR imaging. If a limited number of brain metastases are present, patients often then undergo treatment by radiation oncologists using stereotactic radiosurgery (SRS), a highly precise, targeted treatment to individual metastases that spares normal brain tissue. However, both neuroradiologists and radiation oncologists face a difficult task in detecting and identifying all metastatic disease, particularly small brain metastases that evade the naked eye. Indeed, studies have shown that even expert neuroradiologists fail to detect 25-70% of metastases smaller than 3mm in MR images. Thus, small brain metastases are frequently overlooked, leading to growth, symptom development, and potentially additional treatments.

Computer aided detection (CAD) can potentially augment human performance in this task, as evidenced by studies investigating the improvement in cancer detection in the brain and other sites. However, previously developed CAD systems for brain metastases have exhibited limitations, including a high false positive rate, and especially an inability to consistently detect small metastases. Additionally, the current standard of care in radiation treatment planning and diagnostic imaging review does not incorporate CAD of brain metastases as part of their respective software programs.

Thus, we believe there is a need for 1) an improvement in computer aided detection performance, specifically in the detection of small brain metastases, and 2) the incorporation of this technology into radiation treatment planning and diagnostic imaging software as part of routine clinical practice.

Technology

Duke inventors have developed a novel CAD system that offers superior sensitivity and specificity for detecting brain metastases, outperforming existing algorithms, especially in identifying small metastases that are often overlooked. This is intended to be a tool for more precise diagnosis and treatment planning, ultimately improving patient outcomes for metastatic brain cancer. Specifically, this tool was developed by training and validating an ensemble of convolutional neural networks on a unique dataset containing post-contrast spoiled gradient (SPGR) MRIs of patients with brain metastases, which included many small brain metastases initially overlooked by the treating physician but identified retrospectively (retrospectively identified metastases, or RIMs). The resulting CAD system has been demonstrated to have a sensitivity of 94% for prospectively identified metastases (PIMs), and 80% for RIMs, while maintaining a low false positive rate of 2 per patient, and high segmentation accuracy. For all PIMs and treatable RIMs combined, the size-stratified detection performance in the same study was 100% (> 6 mm), 90% (3-6 mm), and 79% (< 3 mm), superior to prior published results from other studies with a similarly low false positive rate, especially for very small lesions.

Other Applications

This technology could also be applied to routine screening for other types of intracranial tumors, affording improved outcomes by earlier detection.

Advantages

• Superiority: Improvement in small brain metastases detection compared to prior algorithms
• Simplicity: Use of a single MRI sequence
• Sensitivity: 80% of +DC RIMS (not detected and treated prospectively by the initial treating physician but retrospectively meeting diagnostic criteria for positive identification as metastatic disease)
• Specificity: low false positive rate of approximately 3 metastases per 3D MR volume
• High autosegmentation accuracy