SVILUPPO DI UN TOOL PER LA VALUTAZIONE DI ALGORITMI DI SEGMENTAZIONE SU PET

Purpose: cancer is now the second most common cause of death after cardiovascular acute events, but it will soon become the first. Treatment in cases of disseminated disease is often noncurative, toxic and costly. Due to the high variability of response to antineoplastic treatment of different cancers, static and functional imaging play an essential role in therapy efficacy assessment. [18F]fluorodeoxyglucose (FDG) positron emission tomography (PET) has nowadays assumed a key role in oncology for tumor's staging, restaging and follow-up. FDG-PET has always been interpreted qualitatively, comparing areas of high uptake with the surroundings or to reference organs. Recently the quantitative assessment of FDG uptake has been introduced in literature. Besides Standardized Uptake Value (SUV), Metabolic Tumor Volume (MTV) and in particular Total Glycolytic Volume (TGV) have been proposed to be a gauge of tumor burden and aggressiveness. These volumes are calculated using PET segmentation algorithms and written as DICOM-RT structures. A plethora of segmentation methods have been developed but, at now, there is no consensus on which is the best. Before their application in clinical practice there is urgent need for validation, but two problems have to be solved previously. The first is that no realistic ground truth exists that could be used as reference. For that reason we developed a custom phantom containing a tumor insert that has been 3D printed with a radioactive 68Ge gel. The second is that there are no standard metrics to evaluate the best algorithm: several were developed in the past, but they have never been tested on TGV since they all compare the geometrical distance of boundaries and don't take into account the SUV value of voxels. Methods: we extended a C++ software developed to process DICOM files by implementing a function that reads contour points coordinates (mm) of the 3D structure of a DICOM-RT file obtained from a segmentation. Pixels coordinates from spatial points are transformed into a 3D binary matrix (voxels inside the structure are labeled 1, the others are 0). PET matrix with SUV information is then convoluted with mask matrix in order to calculate SUV metrics (SUVmin, SUVmax, SUVmean, SUVrms), MTV and TGV. For the accuracy evaluation we implemented different indexes: the percent difference in volume, the Dice Similarity Coefficient (an index of overlap between the two volumes), the Sensitivity, which gives the portion of test volume correctly classified as tumoral, and the Hausdorff distance, a distance-based method to evaluate the degree of mismatch between two pointsets. These metrics were applied to MTV, and, for the first time, also to TGV. Moreover we developed new metrics that account for both the geometrical properties of volumes and the SUV values inside. To do that we estimated entropies from the histogram of SUV values within segmentations and calculated the mutual information between the RT structures to be assessed and ground truth. We briefly tested the tool on threshold-based segmentation algorithms: percentage and iterative to verify its robustness. Results: the tool has a good automation degree thanks to the use of scripting and provides a complete evaluation of segmentation algorithms. Our preliminary results show that, among all metrics, the DSC and the entropy-based metrics are good for assessing TGV, the first being sensible to boundary variations, the second to the SUV contents of volume.