Abstract
506
Objectives: Intratumor heterogeneity characterized by textural features (TF) from pre-treatment PET images is increasingly used in retrospective studies. Yet, several studies demonstrated that TF are affected by the acquisition and reconstruction parameters, thus complicating multi-center studies and preventing from pooling data acquired in different conditions. We proposed and tested a method to standardized TF measured in PET so that values acquired in different conditions can be compared.
Methods: Patients with breast non-metastatic lesions were included. In department A, 110 patients were scanned before therapy using a Gemini TOF PET/CT scanner (Phillips), with a list-mode iterative reconstruction algorithm (BLOB-OS-TF, 2 iterations, 33 subsets) and a voxel size of 4x4x4 mm3. In department B, 69 patients underwent a pre-treatment PET/CT using a GE Discovery LS Scanner (GMS), reconstructed with OSEM algorithm (2 iterations, 24 subsets) and a voxel size of 2.7x2.7x3.3 mm3. For each patient, VOIs of 23 ml were set in the liver, lung and healthy breast tissue. In each VOI, we computed SUVmean and 6 TF after image quantization using a fixed bin size: Homogeneity, Entropy, SRE, LRE, LGZE and HGZE. For each VOI, we build a vector of biomarkers composed of 6 TF. To characterize the similarity between two biomarker vectors measured in two patients, we computed the ratio between the two vectors (component-wise) and we calculated the standard deviation over the 6 ratios, called SDM as Standard Deviation Metric. If the vectors are homothetic, SDM is 0. We use this metric to measure the similarity between biomarker vectors of different patients. We computed the averaged SDM in patients from dpt A only, in patients from dpt B only and from pairs of patients consisting of 1 from dpt A et 1 from dpt B. We compared the values for TF calculated in three different ways: from the original patient images (M1), from the original patient images resampled to the same voxel size (= 2x2x2 mm3, cubic B-spline interpolation) (M2), and using M2 but then dividing the resulting TF by the mean measured in the liver of that patient (M3).
Results: In the lung tissue, we found that SDM was 0.14±0.09 for patients from dpt A, and 0.19±0.12 for patients from dpt B for the original images (M1), but 0.27±0.15 between patients from dpts A et B. This suggests that the lung TF vector of a patient from dpt A is much more different from the lung TF vector of a patient from dpt B (inter-center variability) than from the lung TF vector of another patient from dpt A (intra-center variability). When calculating TF using M2, SDM was 0.25±0.0.16 between patients from dpts A et B. When normalizing TF as in M3, SDM was 0.12±0.04 between patients from dpts A et B. The same trend was observed in healthy liver and breast tissues.
Conclusion: The proposed standardization method, consisting in first calculating TF from images with the same voxel size and then normalizing the resulting TF by the liver SUVmean, makes the resulting standardized TF comparable between centers using different imaging protocols. Such a standardization could facilitate multicenter study and comparison of TF measured using different imaging protocols. Research Support: ANR-11-IDEX-0003-02