Development of deep neural network-based genotype prediction technology in association with prognosis of brain tumor
Brain glioma is one of the most common primary brain tumors, and glioblastoma multiforme, the most malignant among them, is a very difficult-to-treat disease, which always leads to poor therapeutic outcomes. Among a number of glioma gene mutations, isocitrate dehydrogenase (IDH) genes are considered to be the driver genes of this disease. A large number of studies have revealed that these genes play an important role in determining the resection margin and predicting the treatment response of patients. Therefore, when establishing treatment plans and evaluating clinical trials, it is very important to know whether the mutation of IDH genes has occurred before surgery. To this end, tissues need to be obtained. However, if tumors are located in deep positions, or if patients are aged, there might be risks associated with anesthesia. Thus, determining the presence of IDH gene mutations by collecting patient tissues is a burdensome task.
To overcome this problem, the researchers of the present study attempted to find ways to predict the presence of IDH gene mutations in a non-invasive and reliable manner. Based on the idea that IDH mutations are related to the development of new tumor blood vessels, perfusion-weighted images, by which any changes in blood flow rate in tumors can be monitored, were analyzed using deep learning techniques.
Development of an AI model with high-diagnostic performance that can predict IDH genotypes
The research team took note of the results of previous studies on the correlation between IDH mutations and the development of new tumor blood vessels. In doing so, perfusion-weighted MRI results from brain glioma patients, by which any changes in blood flow rate in tumors can be monitored, were obtained. To analyze these data, a deep learning technique called "convolutional LSTM with attention mechanism" was applied, which was designed to learn the temporal patterns of signals changing over time. The scope of the present study was limited to predicting the presence of IDH mutations through the approach mentioned above and performing a qualitative interpretation as to whether the factors to which the developed deep learning model paid attention were able to represent the nature of such IDH mutations. As a result, it was possible to correctly predict IDH mutations, which would otherwise have been impossible using only existing mean rCBV methods. Also, interestingly, it was found that the model paid attention to different parts depending on the presence of IDH mutations.
Recently, there have been studies using deep learning to predict the presence of IDH mutations, especially based on conventional structural MRI data, such as T2 FLAIR and T1-weighted enhanced images, along with convolutional neural networks. However, too many parameters are involved in the relationship between genotype mutations and the resultant imaging phenotypes of conventional MRI; thus, many agreed that it would be difficult to model this relationship. Crucially, it was also difficult to qualitatively analyze why the deep learning models would make such predictions. In the present study, perfusion-weighted images, which contain more information about IDH mutations, were used for the analysis, and thus it was possible to improve diagnostic performance, such as sensitivity and specificity, and further allow for qualitative analysis, which has significant implications.
The major findings of the present study successfully confirm that overlapped groups can be distinguished with an accuracy of 92.8% by using the primitive T2 * sensitivity signal intensity-time curves obtained by the recurrent neural network model that was designed to learn temporal patterns specific to IDH genotypes out of DSC perfusion MRI data (92.6% of sensitivity and 93.1% of specificity). The developed model determined which temporal aspects were important in predicting IDH genotypes by applying the mentioned mechanism. This approach will make an important contribution to using AI to predict IDH genotypes in a non-invasive manner, which is essential to establishing treatment plans and predicting prognosis in glioma patients, especially those who have high risks associated with anesthesia because they are old, or those whose tumors are located in deep positions.
Prof. Jeong, Bumseok
2019 KI Annual Report