Evaluations of Monte Carlo and AI Techniques for SPECT Reconstruction

Abstract

Molecular radiotherapy is an effective treatment that uses radionuclides bound to molecules that selectively target specific tumor types. SPECT imaging in molecular radiotherapy enables dosimetry, allowing for the estimation of absorbed doses to tumors and normal tissues, which is crucial for optimizing and individualizing treatment. This thesis aims to enhance SPECT imaging and dosimetry in molecular radiotherapy, both in diagnostic imaging before treatment and in imaging after therapy, using Monte Carlo- and AI-based techniques.

Paper I evaluated the spatial resolution in 111In-octreotide SPECT imaging, demonstrating a significant improvement with Monte Carlo-based reconstruction compared to the standard method. Paper II assessed the detectability of simulated liver lesions in 111In-octreotide SPECT images. A visual assessment showed an increase in detection rate from 30.8% to 50.0% using MC-based reconstruction. However, ROC analysis did not show a statistically significant difference between the methods.

Papers III and IV evaluated a convolutional neural network designed to generate synthetic intermediate projections from sparsely acquired SPECT projections. The results demonstrated that acquisition times in 111In-octreotide and 177Lu-DOTATATE SPECT imaging can be substantially reduced, enabling coverage of all organs at risk while maintaining dosimetric accuracy.

For reliable tumor dosimetry, accurate tumor segmentation is essential. In Paper V, a threshold-based method was developed for segmenting liver lesions in SPECT images following molecular radiotherapy with 177Lu-DOTATATE. The method showed promising results for tumors of 4 ml and larger and with tumor-to-normal tissue concentration ratios of 8 or higher.