Graphical User Interface for Viewing Sub-Leisonal Heterogeneity in 18F-NaF PET/CT Fibrous Dysplasia Using Multi-Channel Supervoxel Clustering

Introduction: In traditional nuclear oncology, tumors are often treated as a single entity and summarized into one quantitative value (e.g: SUVmax/mean). Current radiomic analysis techniques and therapeutic development also focus on the tumor as a whole. However, rather than being a single entity, lesions contain internal heterogeneity which can reveal information about regional treatment effectiveness, disease development, and sub-lesional changes. At present, there is a lack of specific and objective tools available to quantify internal tumor heterogeneity. The aim of this abstract is to present an accurate and reproducible method for displaying sub-lesional diversity with an interactive graphical user interface (GUI).

Voxel-level lesion data is difficult to characterize mathematically due to misregistration and long computational times. To interpret the local heterogeneity of a lesion, similar lesional features can be clustered into super-voxels based on location and radiomic features. This way, we can develop a picture of sublesional physiologic characteristics and use textural features to investigate changes in tumor microenvironment after interventions.

Methods: This project used Fibrous Dysplasia (FD) as a model to demonstrate how supervoxel clustering can be used to display lesional heterogeneity. The initial trial evaluated 8 patients with FD, a rare genetic disease resulting in the weakening of the bones. The patients received 120 mg denosumab, which is used to treat FD by increasing bone resorption, every 4 weeks for six months. Patients underwent one baseline and one post-treatment 18F-NaF-PET/CT scan. Lesions were semi-automatically segmented using adaptive thresholding.

The supervoxel clustering algorithm and GUI were developed using MATLAB (version, Mathworks). The initial trial provided voxel-level Hounsfield Unit (HU) and Standard Uptake Value (SUV) data. To capture textural heterogeneity within the lesion, The Image Biomarker Standardization Initiative (IBSI) documentation was used to generate 30 voxel-level radiomic features, including kurtosis, skewness, and GLCM textural values.

Next, we developed the algorithm for clustering voxels together based on feature and textural similarities. The algorithm is based on the idea of "supervoxels", which transforms a lesion into a specified number of clusters based on spatial similarities and image features (see figure).

The GUI is the application of the radiomic feature collection and supervoxel generation algorithm. It allows a user to view the supervoxel parcellation results for a specific lesion using different similarity modes, channels, and viewing options. The application is a simple one-window pop-up, with a straightforward UI that can be utilized without downloading MATLAB. The user uploads their data, and specifies parameters of the lesion of interest and the supervoxel calculation method. The generated image is a video, with the original image data on the left and the transformed data on the right. The data of the specified lesion can be downloaded for analysis.

Results: The result of supervoxel clustering on two characteristic lesions is shown in figure 1. The clustering is determined by the selected textural/radiomic feature.

Conclusions: Our work establishes a user-friendly interface to generate and display heterogeneity within lesions. While this GUI was developed for application in PET/CT FD data, the method of supervoxel generation and display can be generalized for any medical imaging data. We hope this framework can become an accessible tool for researchers and clinicians. Scientists can use this algorithm to determine pre- and post-treatment changes in sub-lesional heterogeneity using statistical measures to evaluate the variance in supervoxel number, size, and texture. Radiologists and clinicians can also use this software to display supervoxel-generated patient data and make qualitative determinations about patient outcomes based on a combination of imaging and textural determinants.

• Download figure
• Open in new tab
• Download powerpoint