Combined Transfer Learning and Test-Time Augmentation Improves Convolutional Neural Network-Based Semantic Segmentation of Prostate Cancer from Multi-Parametric MR Images.

PURPOSE: Multiparametric MRI (mp-MRI) is a widely used tool for diagnosing and staging prostate cancer. The purpose of this study was to evaluate whether transfer learning, unsupervised pre-training and test-time augmentation significantly improved the performance of a convolutional neural network (CNN) for pixel-by-pixel prediction of cancer vs. non-cancer using mp-MRI datasets. METHODS: 154 subjects undergoing mp-MRI were prospectively recruited, 16 of whom subsequently underwent radical prostatectomy. Logistic regression, random forest and CNN models were trained on mp-MRI data using histopathology as the gold standard. Transfer learning, unsupervised pre-training and test-time augmentation were used to boost CNN performance. Models were evaluated using Dice score and area under the receiver operating curve (AUROC) with leave-one-subject-out cross validation. Permutation feature importance testing was performed to evaluate the relative value of each MR contrast to CNN model performance. Statistical significance (p<0.05) was determined using the paired Wilcoxon signed rank test with Benjamini-Hochberg correction for multiple comparisons. RESULTS: Baseline CNN outperformed logistic regression and random forest models. Transfer learning and unsupervised pre-training did not significantly improve CNN performance over baseline; however, test-time augmentation resulted in significantly higher Dice scores over both baseline CNN and CNN plus either of transfer learning or unsupervised pre-training. The best performing model was CNN with transfer learning and test-time augmentation (Dice score of 0.59 and AUROC of 0.93). The most important contrast was apparent diffusion coefficient (ADC), followed by Ktrans and T2, although each contributed significantly to classifier performance. CONCLUSIONS: The addition of transfer learning and test-time augmentation resulted in significant improvement in CNN segmentation performance in a small set of prostate cancer mp-MRI data. Results suggest that these techniques may be more broadly useful for the optimization of deep learning algorithms applied to the problem of semantic segmentation in biomedical image datasets. However, further work is needed to improve the generalizability of the specific model presented herein.

Model-free prostate cancer segmentation from dynamic contrast-enhanced MRI with recurrent convolutional networks: A feasibility study.

Dynamic contrast enhanced (DCE) magnetic resonance imaging (MRI) is a method of temporal imaging that is commonly used to aid in prostate cancer (PCa) diagnosis and staging. Typically, machine learning models designed for the segmentation and detection of PCa will use an engineered scalar image called Ktrans to summarize the information in the DCE time-series images. This work proposes a new model that amalgamates the U-net and the convGRU neural network architectures for the purpose of interpreting DCE time-series in a temporal and spatial basis for segmenting PCa in MR images. Ultimately, experiments show that the proposed model using the DCE time-series images can outperform a baseline U-net segmentation model using Ktrans. However, when other types of scalar MR images are considered by the models, no significant advantage is observed for the proposed model.

Improving fMRI reliability in presurgical mapping for brain tumours.

PURPOSE: Functional MRI (fMRI) is becoming increasingly integrated into clinical practice for presurgical mapping. Current efforts are focused on validating data quality, with reliability being a major factor. In this paper, we demonstrate the utility of a recently developed approach that uses receiver operating characteristic-reliability (ROC-r) to: (1) identify reliable versus unreliable data sets; (2) automatically select processing options to enhance data quality; and (3) automatically select individualised thresholds for activation maps. METHODS: Presurgical fMRI was conducted in 16 patients undergoing surgical treatment for brain tumours. Within-session test-retest fMRI was conducted, and ROC-reliability of the patient group was compared to a previous healthy control cohort. Individually optimised preprocessing pipelines were determined to improve reliability. Spatial correspondence was assessed by comparing the fMRI results to intraoperative cortical stimulation mapping, in terms of the distance to the nearest active fMRI voxel. RESULTS: The average ROC-r reliability for the patients was 0.58±0.03, as compared to 0.72±0.02 in healthy controls. For the patient group, this increased significantly to 0.65±0.02 by adopting optimised preprocessing pipelines. Co-localisation of the fMRI maps with cortical stimulation was significantly better for more reliable versus less reliable data sets (8.3±0.9 vs 29±3 mm, respectively). CONCLUSIONS: We demonstrated ROC-r analysis for identifying reliable fMRI data sets, choosing optimal postprocessing pipelines, and selecting patient-specific thresholds. Data sets with higher reliability also showed closer spatial correspondence to cortical stimulation. ROC-r can thus identify poor fMRI data at time of scanning, allowing for repeat scans when necessary. ROC-r analysis provides optimised and automated fMRI processing for improved presurgical mapping.

Enhancement of bone consolidation in mandibular distraction osteogenesis: a contemporary review of experimental studies involving adjuvant therapies.

BACKGROUND: One of the major disadvantages of mandibular distraction osteogenesis (MDO) is the prolonged time required for consolidation of the regenerate bone. The objective of the present study is to perform a contemporary review of various adjuvant therapies to enhance bone consolidation in MDO. METHODS: A PubMed search for articles related to MDO, along with the references of those articles, was performed. Inclusion and exclusion criteria were applied to all experimental studies assessing adjuvant therapies to enhance bone consolidation. RESULTS: A total of 1414 titles and abstracts were initially reviewed; 61 studies were included for full review. Many studies involved growth factors, hormones, pharmacological agents, gene therapy, and stem cells. Other adjuvant therapies included mechanical stimulation, laser therapy, and hyperbaric oxygen. Majority of the studies demonstrated positive bone healing effects and thus adjuvant therapies remain a viable strategy to enhance and hasten the consolidation period. CONCLUSION: Although most studies have demonstrated promising results, many questions still remain, such as optimal amount, timing, and delivery methods required to stimulate the most favorable bone regeneration. As well, further studies comparing various adjuvant therapies and documentation of long-term adverse effects are required prior to clinical application.

Quantification of superparamagnetic iron oxide with large dynamic range using TurboSPI.

This work proposes the use of TurboSPI, a multi-echo single point imaging sequence, for the quantification of labeled cells containing moderate to high concentrations of iron oxide contrast agent. At each k-space location, TurboSPI acquires several hundred time points during a spin echo, permitting reliable relaxation rate mapping of large-R(2)(∗) materials. An automatic calibration routine optimizes image quality by promoting coherent alignment of spin and stimulated echoes throughout the multi-echo train, and this calibration is sufficiently robust for in vivo applications. In vitro relaxation rate measurements of SPIO-loaded cervical cancer cells exhibit behavior consistent with theoretical predictions of the static dephasing regime in the spin echo case; the relaxivity measured with TurboSPI was 10.47±2.3 s(-1)/mG, comparable to the theoretical value of 10.78 s(-1)/mG. Similar measurements of micron-sized iron oxide particles (0.96 µm and 1.63 µm diameter) show a reduced relaxivity of 8.06±0.68 s(-1)/mG and 7.13±0.31 s(-1)/mG respectively, indicating that the static dephasing criterion was not met. Nonetheless, accurate quantification of such particles is demonstrated up to R(2)(∗)=900 s(-1), with a potentially higher upper limit for loaded cells having a more favorable R(2)('):R(2) ratio. Based on the cells used in this study, reliable quantification of cells loaded with 10 pg of iron per cell should be possible up to a density of 27 million cells/mL. Such quantification will be of crucial importance to the development of longitudinal monitoring for cellular therapy and other procedures using iron-labeled cells.

Measurement of fluid ingress into calcium polyphosphate bioceramics using nuclear magnetic resonance microscopy.

A MR microscopy experiment is developed and used to characterize fluid ingress and microstructural transformation in degradable calcium polyphosphate (CPP) bioceramics. High-resolution (49microm) maps of fluid density and spin-lattice relaxation rate were obtained as a function of time for CPP immersed in phosphate buffered saline. These results demonstrate clear differences in fluid transport rates and solid matrix microstructure in two differing CPP formulations. CPP has been proposed as a potential implantable device for the delivery of pharmaceuticals, and the magnetic resonance imaging (MRI) data are used in conjunction with previously reported bulk elution results to develop a hypothesis explaining microstructural evolution in these materials. This type of non-destructive evaluation of the structure-transport of fluids in CPP is important to improved design of these functionalized biomaterials for long-term, localized delivery of sustained levels of therapeutic agents.

Studies of porous media by thermally polarized gas NMR: current status.

Three examples of thermally polarized gas NMR performed at New Mexico Resonance are presented to demonstrate its unique advantages in porous media studies. 1) In-vivo animal lung imaging by Kuethe et al., in which useful quality 3D images of rat lungs were obtained in 30 min. It is conjectured that comparable human lung images would take much less time to make, possibly by the ratio of body weights, a factor of several hundred. 2) The success of the lung imaging suggested other porous media as candidates for thermally polarized gas NMR. Caprihan and coworkers obtained excellent images from partially sintered ceramics and Vycor glass. Since then, Beyea has developed the technique of spatially resolved BET curves for ceramics and other nanoporous solids. In this way, surface area, pore size, and porosity, averaged over an image voxel, can be spatially resolved. This greatly aids in the characterization of such materials, especially with regards to spatial heterogeneities. 3) Finally, we describe Codd's propagator experiments on propane gas flowing through a packed bed of 300 microm beads. In order to increase signal-to-noise ratio, the flowing gas was pressurized to 170 kPa. Excellent quality propagators showing the discrete nature of the bead pack were obtained. This type of information is not available in comparable liquid studies because most spins will not diffuse far enough to sample the walls in the time available.