Abstract
T41
Introduction: With the expanding field of radiopharmaceutical therapies (RPT) and interest in enabling personalized cancer treatments, there is still work to do in streamlining the patient-specific dosimetry workflow. Segmentation of the organs and tumors, which is a laborious process, has historically been performed by the medical physicists with the nuclear medicine physician (NMP) reviewing these volumes of interest (VOI). With the inclusion of artificial intelligence (AI) algorithms trained for segmentation in different imaging software, this process has become faster and automated. Moreover, these new tools provide a new opportunity to include the nuclear medicine technologist (NMT) in the segmentation process of the dosimetry workflow. The aim of this study was to train a NMT in the segmentation for dosimetry purposes of patients treated with 177Lu-PSMA-617 and to evaluate a commercial AI segmentation algorithm.
Methods: A workflow to enable easy segmentation of organs and tumors in 177Lu-PSMA-617 SPECT/CT image was built using MIM (MIM Software, USA). A total of 73 imaging time points from 15 patients were analyzed. A qPSMA approach using the liver as reference, was used on the quantitative SPECTs to define lesions. Because dosimetry requires the knowledge of activity and mass, we estimated the total tumor burden volume (TTBv) and activity (TTBa). To determine the mass of organs (left kidney (LKv), right kidney (RKv), liver (Lv), and spleen (Sv)), contouring was performed using the MIM AI algorithm on the CT images. Lv and Sv regions were transferred to the SPECT to determine liver (La) and spleen (Sa) activities. The activity of the left kidney (LKa), right kidney (RKa), and salivary glands (SGa) was measured using a 40% threshold manually performed by the NMT. The NMT reviewed all the AI VOIs created on the CT. All the VOIs were then verified and approved or modified accordingly by a NMP. The % difference in change of organ volumes and activities between AI to NMT and NMT to NMP were determined for LKv,a, RKv,a, Lv,a, Sv,a, TTBv,a. A paired t-test was used to identify if differences between VOIs from AI with NMT and NMT with NMP were statistically significant.
Results: From the 73 images, 9 (LKv, RKv), 25 (Lv,La) and 33 (Sv,Sa) adjustments were made by the NMT to the AI VOIs. The NMP modified 3 (LKv), 2 (LKa), 7 (TTBv,a) from the NMT accepted VOIs. No other changes were required including SGa. The % difference in change for VOI measured parameters between AI and NMT were LKv (2.88±1.45), RKv (31.25±38.56), Lv (0.09±0.65), Sv (-4.17±7.07), La (1.12±2.54), and Sa (-0.88±11.73). Similarly, % differences between the NMT and NMP were only found for LKv (4.89±6.31), LKa (11.01±5.17), TTBv (-16.51±31.74), and TTBa (-15.79±29.10). The only statistically significant difference was observed for Sv (p=0.0035), primarily due to AI not recognizing the anterior and posterior splenic extremities. Large difference for RKv was influenced by renal cysts. Differences in TTBv,a was due to bowel encroachment or missed lesions.
Conclusions: We streamlined the segmentation process for 177Lu-PSMA-617 dosimetry by involving the NMT and using AI algorithms. The use of AI was validated by comparing the organ and tumor contours to the NMT, with confirmation by the NMP. Most of the AI segmented VOIs did not require any modifications. The few that required them were due to anatomy variances (e.g. cyst). However, the differences in activity and volumes required for dosimetry were not statistically significant. Our results suggest that AI is a reliable tool for segmentation for dosimetry purposes and that NMTs can add value for practical, routine dosimetry implementation in the clinic.