Search

Background: Prostate‐specific membrane antigen (PSMA) PET imaging represents a valuable source of information reflecting disease stage, response rate, and treatment optimization options, particularly with PSMA radioligand therapy. Quantification of radiopharmaceutical uptake in healthy organs from PSMA images has the potential to minimize toxicity by extrapolation of the radiation dose delivery towards personalization of therapy. However, segmentation and quantification of uptake in organs requires labor‐intensive organ delineations that are often not feasible in the clinic nor scalable for large clinical trials. Purpose: In this work we develop and test the PSMA Healthy organ segmentation network (PSMA‐Hornet), a fully‐automated deep neural net for simultaneous segmentation of 14 healthy organs representing the normal biodistribution of [18F]DCFPyL on PET/CT images. We also propose a modified U‐net architecture, a self‐supervised pre‐training method for PET/CT images, a multi‐target Dice loss, and multi‐target batch balancing to effectively train PSMA‐Hornet and similar networks. Methods: The study used manually‐segmented [18F]DCFPyL PET/CT images from 100 subjects, and 526 similar images without segmentations. The unsegmented images were used for self‐supervised model pretraining. For supervised training, Monte‐Carlo cross‐validation was used to evaluate the network performance, with 85 subjects in each trial reserved for model training, 5 for validation, and 10 for testing. Image segmentation and quantification metrics were evaluated on the test folds with respect to manual segmentations by a nuclear medicine physician, and compared to inter‐rater agreement. The model's segmentation performance was also evaluated on a separate set of 19 images with high tumor load. Results: With our best model, the lowest mean Dice coefficient on the test set was 0.826 for the sublingual gland, and the highest was 0.964 for liver. The highest mean error in tracer uptake quantification was 13.9% in the sublingual gland. Self‐supervised pretraining improved training convergence, train‐to‐test generalization, and segmentation quality. In addition, we found that a multi‐target network produced significantly higher segmentation accuracy than single‐organ networks. Conclusions: The developed network can be used to automatically obtain high‐quality organ segmentations for PSMA image analysis tasks. It can be used to reproducibly extract imaging data, and holds promise for clinical applications such as personalized radiation dosimetry and improved radioligand therapy. [ABSTRACT FROM AUTHOR]

Application of kinetic modeling (KM) on a voxel level in dynamic PET images frequently suffers from high levels of noise, drastically reducing the precision of parametric image analysis. In this paper, we investigate the use of machine learning and artificial neural networks to denoise dynamic PET images. We train a deep denoising autoencoder (DAE) using noisy and noise-free spatiotemporal image patches, extracted from the simulated images of [11C]raclopride, a dopamine D2 receptor agonist. The DAE-processed dynamic and corresponding parametric images (simulated and acquired) are compared with those obtained with conventional denoising techniques, including temporal and spatial Gaussian smoothing, iterative spatiotemporal smoothing/deconvolution, and the highly constrained backprojection processing (HYPR). The simulated (acquired) parametric image non-uniformity was 7.75% (19.49%) with temporal and 5.90% (14.50%) with spatial smoothing, 5.82% (16.21%) with smoothing/deconvolution, 5.49% (13.38%) with HYPR, and 3.52% (11.41%) with DAE. The DAE also produced the best results in terms of the coefficient of variation of voxel values and structural similarity index. Denoising-induced bias in the regional mean binding potential was 7.8% with temporal and 26.31% with spatial smoothing, 28.61% with smoothing/deconvolution, 27.63% with HYPR, and 14.8% with DAE. When the test data did not match the training data, erroneous outcomes were obtained. Our results demonstrate that a deep DAE can provide a substantial reduction in the voxel-level noise compared with the conventional spatiotemporal denoising methods while introducing a similar or lower amount of bias. The better DAE performance comes at the cost of lower generality and requiring appropriate training data. [ABSTRACT FROM AUTHOR]

• We tested the ability of convolutional neural nets to effectively capture intensity, shape, and texture of tumors in medical images. • Convolutional neural networks were found to have a limited capacity to capture tumor shape and heterogeneity. • Shape and texture features required an order-of-magnitude more training samples compared to intensity features. • Convolutional neural networks combined with explicit radiomics should yield the most powerful predictive models. Radiomics and deep learning have emerged as two distinct approaches to medical image analysis. However, their relative expressive power remains largely unknown. Theoretically, hand-crafted radiomic features represent a mere subset of features that neural networks can approximate, thus making deep learning a more powerful approach. On the other hand, automated learning of hand-crafted features may require a prohibitively large number of training samples. Here we directly test the ability of convolutional neural networks (CNNs) to learn and predict the intensity, shape, and texture properties of tumors as defined by standardized radiomic features. Conventional 2D and 3D CNN architectures with an increasing number of convolutional layers were trained to predict the values of 16 standardized radiomic features from real and synthetic PET images of tumors, and tested. In addition, several ImageNet-pretrained advanced networks were tested. A total of 4000 images were used for training, 500 for validation, and 500 for testing. Features quantifying size and intensity were predicted with high accuracy, while shape irregularity and heterogeneity features had very high prediction errors and generalized poorly. For example, mean normalized prediction error of tumor diameter with a 5-layer CNN was 4.23 ± 0.25, while the error for tumor sphericity was 15.64 ± 0.93. We additionally found that learning shape features required an order of magnitude more samples compared to intensity and size features. Our findings imply that CNNs trained to perform various image-based clinical tasks may generally under-utilize the shape and texture information that is more easily captured by radiomics. We speculate that to improve the CNN performance, shape and texture features can be computed explicitly and added as auxiliary variables to the networks, or supplied as synthetic inputs. [ABSTRACT FROM AUTHOR]