Dr. Minjeong Kim (Computer Science) received new funding from Cone Health System for the project “AI Assisted Early Outcome Predication for Brain Tumors Treated with Stereotactic Radiosurgery.”
Nearly 2 million people in the United States are expected to be diagnosed with cancer each year. Of those 2 million people, up to 40% will develop brain metastases. Stereotactic radiosurgery (SRS) is a popular choice to treat brain metastases due to its non-invasive nature, convenience to the patient, and a reduced risk of cognitive impairment compared to other treatment techniques. However, SRS does have limitations, including the risk of distant recurrence and radionecrosis, which necessitates active monitoring and surveillance scanning with magnetic resonance (MR) imaging.
The monitoring of the patient is complicated challenging due to the difficulty associated with differentiating between tumor progression, radionecrosis, and transient treatment -induced effects. The only sure way to evaluate the tumor status is through histology, which is an minimally invasive process that requires a tissue sample. Tissue samples, that may not even be possible which is in some cases not possible due to tumor location.
The current standard of care in these cases may involve multiple types of imaging (different types of MR imaging and positron emission tomography) and observation over time. This process can be substantially is difficult and complicated for the patients because of the multiple appointments and time spent waiting, and costly depending on the number of clinic visits and cost needed, and it may. This process also delays the initiation of care for progressing disease or radionecrosis. In addition, evaluation of tumor status typically involves manual measurements of tumor geometry, which are labor labor-intensive and subject to inter- and intra-observer variability. The volume of follow up images scans in a busy clinic may place a marked burden of time and effort on the radiation oncology and radiology staff because of the time and effort.
The researchers approach computer-assisted diagnosis (CAD) as a solution for addressing the clinical challenges mentioned above. By applying an automated software tool for diagnosing and predicting tumor, the researchers can relieve the social burden, the cost to the patients, as well as the health system, and save time for the radiologists/radiation oncologists by providing decision-supportive information in advance to follow-up visits and future diagnosis. Specifically, the known information of tumor progression and the patient outcome, along with the corresponding imaging data, collected retroactively can serve as training data for the diagnostic model.
As with the recent advancement of artificial intelligence (in particular, machine/deep learning), the researchers can design and build a machine/deep learning-based diagnostic regression model that predicts disease progression and outcome at an earlier time for individual patients.
The model is to learn the relationship between the intrinsic data representations of longitudinal MR images and their respective diagnostic information (often using data not easily processed by humans), including outcome, progression, radionecrosis, and treatment-related factors, as the target values by the optimization process to maximize the prediction accuracy. The success of achieving high prediction accuracy will allow for earlier interventions that can be more accurate and consistent than human experts. A diagnostic tool that could provide definitive information earlier in the process would reduce the social burden on the patient, reduce the cost to the patient and health system, save time for the radiologists, radiation oncologists, and allow for earlier interventions.
Computer assisted diagnosis (CAD) offers the possibility of an earlier, definitive diagnosis without the need for a biopsy and watchful waiting. This would greatly reduce the burden on the patient and staff, and potentially allow for earlier intervention. Computer assisted diagnosis leverages machine learning to help extract features of images that a human might not be able to see and uses those to determine a diagnosis or outcome.
In this case, the images of interest would be the serial MR images that are acquired for monitoring. The CAD would be able to predict probabilities of progression, radionecrosis, and treatment related effects based on analysis of the images, hopefully at a point earlier than a human could. The CAD would be trained using a retrospective data set from previously treated patients with known outcomes.
