Deep Learning-Based Partial Volume Correction in Lu-177 SPECT/CT Imaging

Abstract

This research internship investigated whether a deep learning (DL) model can
compensate for partial volume effects in Lu-177 single-photon emission computed tomography (SPECT) images. The project focused on a reproducible
pipeline for training and evaluating residual 2.5D U-shaped convolutional neural network (U-Net) models that predict corrected activity distributions from
low-resolution reconstructions. A central part of the work was lesion-level
evaluation, because clinically relevant performance depends not only on image similarity but also on recovered lesion activity.
The implemented workflow includes patient-level train/validation/test splitting,
lesion-aware sampling, residual prediction, activity-weighted losses, final holdout evaluation, and analysis of lesion activity recovery. A separate resampling
study quantified the information loss introduced when 512 matrix lesion segmentations are represented in 128 matrix space. On the calibrated final test
cohort, the model increased the mean lesion recovery coefficient (RC) from
0.363 for the forward-projected reconstruction to 0.586, and reduced mean
relative lesion activity error from 0.679 to 0.542 across 388 lesions. The
results show that deep learning-based partial volume correction (PVC) can
recover a substantial part of the missing lesion activity, while also revealing remaining challenges in calibration, small-lesion stability, and timepointdependent performance.