0023

Predictive and discriminative localization of IDH genotype in high grade gliomas using deep convolutional neural nets
Adnan Ahmad1, Srinjay Sarkar1, Apurva Shah1, Santosh Vani2, Jitender Saini3, and Madhura Ingalhalikar1

1Symbiosis Centre of Medical Image Analysis, Symbiosis International University, Pune, India, 2National Institute of Mental Health and Neurosciences, Bangalore, India, 3Department of Neuroimaging & Interventional Radiology, National Institute of Mental Health and Neurosciences, Bangalore, India

Synopsis

Radiomics and state-of-art convolutional neural networks (CNNs) have demonstrated their usefulness for predicting genotype in gliomas from brain MRI. However, these techniques rely on accurate tumor segmentation and do not facilitate insights into the critical discrimative features. To mitigate this, we employ a novel technique called CNNs with discriminative localization (DL-CNN) on a clinical T2 weighted MRI dataset of IDH1 mutant and wild-type tumor patients, which is not only free of tumor segmentation with high classification accuracy of 86.7% but also demonstrates that the tumoral area is discriminative in mutants while in IDH1 wildtype the peri-tumoral edema is also involved.

Introduction

Recent WHO classification of brain tumors has introduced genomic characterization of gliomas as studies have demonstrated that mutations in isocitrate dehydrogenase 1(IDH1) are associated with longer overall survival1. Currently, the IDH genotype is identified via immuno-histochemical analysis following biopsy or surgical resection. Thus, developing a non-invasive technique for pre-operative prediction of the IDH status is critical and clinically significant, as it can enable patient-specific treatment plan as well as support therapeutic intervention. Existing techniques using radiomics and convolutional neural nets (CNNs) employ multi-modal tumor segmentation as a pre-processing step, which is complicated and time-consuming. Furthermore, despite the accurate performance of CNNs, gaining clinical interpretability from the CNN is imperative as although the model achieves a good classification performance, it could be susceptible to over-fitting especially on smaller sample sizes. This work proposes a novel generalized technique called CNN based discriminative localization (DL-CNN) which (1) eliminates the need of tumor segmentation by involving a much larger region of interest and (2) employs a deep network and global average pooling to reverse map the most discriminative regions on the input image to create a class activation map (CAM). The proposed framework is highly relevant in a clinical setting as the CAMs provide clinically meaningful insights into the tumor regions.

Methods

Our dataset was collected at the National Institute of Mental Health and Neurological Sciences (NIMHANS), Bangalore. IDH1 genotype was identified in 41 subjects’ (40.26 ± 12.09 years, 26M/15F) while 30 subjects (46.8 ± 18.86 years, 14M/16F) were wild-type. Image acquisition was performed on Siemens 3T Skyra and Philips 3T Achieva. T2 scans were acquired using a TR/TE ranging from 3600-5500/80-90 ms and 1*1 mm resolution in the axial plane. Our model employed only the axial slices (1mm*1mm resolution) that presented with the tumor. To harmonize the data between scanners, we employed normalization before the initial convolutional layer. To classify the data we employed a Deep Convolutional Architecture based on the Resnet-50 framework (Figure 1). While training, the model converged on the categorical cross entropy loss using the Adam optimization function with a decaying learning rate of 10e-4. In addition, to generate the CAMs we isolated the outputs from the last Convolutional layer of the model for all the cases. We used these activations as inputs to a new model as shown in Figure 1 and extracted the activations of its convolutional layer and upsampled it to match the input image2. To compare model performance we created 3 instances of each image: 1. manually, segmented tumor inputs, 2. non-segmented data, but with a boxed region of interest and 3. on complete axial slices.

Results

Our ResNet architecture with boxed area of interest was able to achieve an accuracy of 80% (4 fold-cross validation) and a test accuracy of 86.7%, while when full axial slices were employed the accuracy dropped down to 75.05% (4-fold cross validation) and 83.5% (testing). In case of manually segmented tumors the accuracy further dropped to 71.3% (4-fold cross validation) and 72.73% (testing). The receiver operating curves (ROC) for the three cases were also computed (figure 2). We display the CAMs in figure 3 for a few test subjects that were correctly classified. The first row of the color maps shows the IDH1 mutants while the second row shows the wild-types.

Discussion

The method implemented here not only performs prediction but also provides clinical interpretability by creating DL-CNNs and applying it to an extremely pertinent problem of genomic sub-typing of gliomas. We demonstrate that the tumor segmentation can be completely eliminated while the high activation maps or CAMs vary distinctly between the two classes. IDH1 wildtype gliomas demonstrate the involvement of peri-tumoral edema in addition to the tumor core, which can be attributed to their highly infiltrative nature3 in comparison to the mutant cases. Imaging prediction of IDH mutation could be crucial in future as IDH mutant inhibitors become clinically available, these might be used as neoadjuvant therapy4.

Conclusion

We demonstrate that DL-CNNs can classify high grade gliomas with IDH1 mutation from the wild-type without segmenting the tumor. Patient specific CAMs highlighted the tumoral area in mutants while in the wildtype cases the peri-tumoral edema was also significant. A computer aided framework can be built to attain IDH1 predictions as well as to gain insights into the areas of significance that could support targeted therapy.

Acknowledgements

No acknowledgement found.

References

  1. C. Houillier, X. Wang, G. Kaloshi, K. Mokhtari, R. Guillevin, J. Laffaire, et al., "IDH1 or IDH2 mutations predict longer survival and response to temozolomide in low-grade gliomas," Neurology, vol. 75, pp. 1560-6, Oct 26 2010.
  2. B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, "Learning Deep Features for Discriminative Localization," CVPR, 2016.
  3. S. J. Price, K. Allinson, H. Liu, N. R. Boonzaier, J. L. Yan, V. C. Lupson, et al., "Less Invasive Phenotype Found in Isocitrate Dehydrogenase-mutated Glioblastomas than in Isocitrate Dehydrogenase Wild-Type Glioblastomas: A Diffusion-Tensor Imaging Study," Radiology, vol. 283, pp. 215-221, Apr 2017.
  4. L. Dang, K. Yen, and E. C. Attar, "IDH mutations in cancer and progress toward development of targeted therapeutics," Ann Oncol, vol. 27, pp. 599-608, Apr 2016.

Figures

Figure 1:Schematic diagram of the DL-CNN model. 2D images are fed to the ResNet with 5 stages. Each stage consists of a conv. block and 3 identity blocks. The structure of identity block is shown with the short cut connection. The output feature map of this CNN is G which once learned from the prediction model, is passed on to create the CAMs as shown.

Figure 2:ROC of ResNet

Figure 3:Figure showing the input test images and their color maps. These color maps indicate the most discriminative regions of the image on their right. (a-c: upper row) are IDH1 mutant cases while (d-f: lower row) are IDH1 wild-type. The red color indicates the most elevated activations followed by yellow and cyan.

Table 1: Comparative results

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
0023