5526. Exploring Interpretation Maps as a Path to Discover Radiogenomics Biomarkers: A Call for Rethinking
Authors* Denotes Presenting Author
Shahriar Faghani;
Mayo Clinic - Rochester
Mana Moassefi *;
Mayo Clinic - Rochester
Objective:
Deep learning (DL) has shown promising results in predicting genetic status from imaging. While these models can accurately predict genetic profiles from imaging, they often lack interpretability. Integrated gradients (IG) is a technique that assigns importance scores to pixels based on their impact on model output. IG maps provide insight into the model’s decision-making process and the contribution of each pixel to the output. This study examines the role of IG maps in providing interpretation on IDH and MGMT classification in glioblastomas using MRI.
Materials and Methods:
We utilized the publicly available UCSF-PDGM dataset to analyze various MRI sequences along with corresponding tumor genetic profiles. Our study used these sequences independently and in combination to train 3D-Densenet121 models that could predict the IDH mutation and MGMT promoter methylation status. We used 3D IG maps to identify imaging areas that influenced decision-making.
Results:
The area under the receiver operating curve (AUC) for IDH classification were 0.94, 0.93, and 0.94 on T2, contrast-enhanced T1 (CT1), and T2-CT1, respectively. Our trained model on T2 for MGMT prediction achieved an AUC of 0.65. For IDH classification on T2 the model primarily highlighted cerebrospinal fluid (CSF), whereas on CT1, it focused mainly on contrast-enhanced areas. However, when both CT1 and T2 images were used, the model focused primarily on the tumoral area. In contrast, the model for the MGMT task highlighted all areas without any specific anatomical or functional preference.
Conclusion:
In radiogenomics, the imaging features that a model detects, such as IDH classification, depend on the imaging sequence. By using multiparametric MRI, the model's attention is more directed toward the tumor region. However, if the model is not able to accurately predict the outcome of the radiogenomics task, there are no particular imaging areas that are emphasized. To draw a general conclusion, it is necessary to use more interpretation maps, employ different combinations of MRI sequences, and analyze the reports of neuroradiologists on the highlighted areas. This approach introduces a novel method to identify imaging biomarkers for radiogenomics via DL.