PT  - JOURNAL ARTICLE
AU  - Anne-Sophie Dirand
AU  - Fanny Orlhac
AU  - Christophe Nioche
AU  - Irene Buvat
TI  - Voxel-based analysis facilitates the identification of robust machine-learning based models for tumor classification
DP  - 2020 May 01
TA  - Journal of Nuclear Medicine
PG  - 271--271
VI  - 61
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/61/supplement_1/271.short
4100  - http://jnm.snmjournals.org/content/61/supplement_1/271.full
SO  - J Nucl Med2020 May 01; 61
AB  - 271Objectives: Designing robust predictive or prognostic radiomic models using machine-learning approaches is often challenging due to the limited number of patients available for training. Using a voxel-based modelling approach instead of a patient-based modelling considerably increases the number of observations as there are frequently dozens to hundreds of voxels in the region of interest of a single patient. We therefore designed an original voxel-based modelling approach and studied whether it could facilitate the identification of robust radiomic models compared to a region-based approach. Methods: A previously published cohort of 48 soft tissue sarcoma patients including PET data [1] was used. The task was to predict the metastasis status of each patient from radiomic features measured in the main tumor. Each patient was labeled either S0 (no lung metastasis) or S1 (with lung metastases) based on a biopsy or diagnosed by an expert physician. For each patient, 5 parametric images of Dissimilarity, Contrast, Energy, Entropy and Homogeneity were calculated using LIFEx [2] with 3 different sizes of kernel, yielding 15 feature values per voxel. The data set was then separated into a dictionary set (DS) of 29 patients and a test set (TS) of 19 patients. Using the DS, each of the 15 feature values composing the voxel profile was scaled between 0 and 1 and these scaling coefficients were applied to the TS voxel profiles. For each patient, a principal component (PC) analysis of all voxel profiles within a region of interest drawn manually over the main tumor (ROIT) was performed and the first n PC needed to recover 99% of the explained variance were selected, corresponding to n representative “supervoxel” profiles describing the tumor. For each patient k of TS, each of its selected nk PC was paired with the closest PC calculated from the DS using the Euclidean distance and the status of the patient associated with that closest PC was stored. Once all nk PCs were paired, this patient was described as having NS1% of S1 PC. After the k patients of TS were studied, a ROC analysis was performed to determine whether the NS1 percentages could distinguish between the S1 and S0 patient status. This process was conducted 100 times by randomly splitting the DS and TS to study the robustness and reliability of the results. Last, to compare the voxel-based model performance with the performance of a region-based approach, the 5 radiomic features used for calculating the parametric images were calculated over ROIT in all patients. Each TS patient tumor profile was thus defined as the vector of the 5 radiomic feature values calculated over ROIT. It was then compared to the 5-feature tumor profiles of the patients in the DS using a weighted Euclidean distance. The TS patient status was assigned the status of the DS patient who had the closest profile. Results: Using the voxel-based analysis, the area under the ROC curve (AUC±1sd) to distinguish between the S0 and S1 status in the VS was 0.79±0.11, corresponding to a Youden index (sensitivity+specificity-1) ±1sd of 0.58±0.17. With the region-based approach (one profile per patient), the Youden index was 0.37±0.16, significantly lower than when using the voxel-based approach (p<0.05). Conclusions: Using the exact same data and 5 radiomic features only, the voxel-based method yielded a model with significantly better classification performance than the region-based method. These results suggest that a voxel-based analysis of tumors might facilitate the design of robust radiomic models using machine learning approaches when the number of patients is limited.