Comparison of quantitative parameters and radiomic features as inputs into machine learning models to predict the Gleason score of prostate cancer lesions.

INTRODUCTION: The classification of prostate cancer (PCa) lesions using Prostate Imaging Reporting and Data System (PI-RADS) suffers from poor inter-reader agreement. This study compared quantitative parameters or radiomic features from multiparametric magnetic resonance imaging (mpMRI) or positron emission tomography (PET), as inputs into machine learning (ML) to predict the Gleason scores (GS) of detected lesions for improved PCa lesion classification. METHODS: 20 biopsy-confirmed PCa subjects underwent imaging before radical prostatectomy. A pathologist assigned GS from tumour tissue. Two radiologists and one nuclear medicine physician delineated the lesions on the mpMR and PET images, yielding 45 lesion inputs. Seven quantitative parameters were extracted from the lesions, namely T2-weighted (T2w) image intensity, apparent diffusion coefficient (ADC), transfer constant (KTRANS), efflux rate constant (Kep), and extracellular volume ratio (Ve) from mpMR images, and SUVmean and SUVmax from PET images. Eight radiomic features were selected out of 109 radiomic features from T2w, ADC and PET images. Quantitative parameters or radiomic features, with risk factors of age, prostate-specific antigen (PSA), PSA density and volume, of 45 different lesion inputs were input in different combinations into four ML models - Decision Tree (DT), Support Vector Machine (SVM), k-Nearest-Neighbour (kNN), Ensembles model (EM). RESULTS: SUVmax yielded the highest accuracy in discriminating detected lesions. Among the 4 ML models, kNN yielded the highest accuracies of 0.929 using either quantitative parameters or radiomic features with risk factors as input. CONCLUSIONS: ML models' performance is dependent on the input combinations and risk factors further improve ML classification accuracy.

Quantitative 68Ga-DOTATATE PET/CT Parameters for the Prediction of Therapy Response in Patients with Progressive Metastatic Neuroendocrine Tumors Treated with 177Lu-DOTATATE.

The aim of this study was to determine whether quantitative PET parameters on baseline 68Ga-DOTATATE PET/CT and interim PET (iPET) performed before the second cycle of therapy are predictive of the therapy response and progression-free survival (PFS). Methods: Ninety-one patients with well-differentiated neuroendocrine tumors (mean Ki-67 index, 8.3%) underwent 68Ga-DOTATATE PET/CT to determine suitability for peptide receptor radionuclide therapy as part of a prospective multicenter study. The mean follow-up was 12.2 mo. Of the 91 patients, 36 had iPET. The tumor metrics evaluated were marker lesion-based measures (mean SUVmax and ratio of the mean lesion SUVmax to the SUVmax in the liver or the SUVmax in the spleen), segmented 68Ga-DOTATATE tumor volumes (DTTVs), SUVmax and SUVmean obtained with the liver and spleen as thresholds, and heterogeneity parameters (coefficient of variation, kurtosis, and skewness). The Wilcoxon rank sum test was used for the association between continuous variables and the therapy response, as determined by the clinical response. Univariable and multivariable Cox proportional hazards models were used for the association with PFS. Results: There were 71 responders and 20 nonresponders. When marker lesions were used, higher mean SUVmax and ratio of the mean lesion SUVmax to the SUVmax in the liver were predictors of the therapy response (P = 0.018 and 0.024, respectively). For DTTV parameters, higher SUVmax and SUVmean obtained with the liver as a threshold and lower kurtosis were predictors of a favorable response (P = 0.025, 0.0055, and 0.031, respectively). The latter also correlated with a longer PFS. The iPET DTTV SUVmean obtained with the liver as a threshold and the ratio of mean SUVmax obtained from target lesions at iPET to baseline PET correlated with the therapy response (P = 0.024 and 0.048, respectively) but not PFS. From the multivariable analysis with adjustment for age, primary site, and Ki-67 index, the mean SUVmax (P = 0.019), ratio of the mean lesion SUVmax to the SUVmax in the liver (P = 0.018), ratio of the mean lesion SUVmax to the SUVmax in the spleen (P = 0.041), DTTV SUVmean obtained with the liver (P = 0.0052), and skewness (P = 0.048) remained significant predictors of PFS. Conclusion: The degree of somatostatin receptor expression and tumor heterogeneity, as represented by several metrics in our analysis, were predictive of the therapy response or PFS. Changes in these parameters after the first cycle of peptide receptor radionuclide therapy did not correlate with clinical outcomes.

Validation of low-dose lung cancer PET-CT protocol and PET image improvement using machine learning.

PURPOSE: To conduct a simplified lesion-detection task of a low-dose (LD) PET-CT protocol for frequent lung screening using 30% of the effective PETCT dose and to investigate the feasibility of increasing clinical value of low-statistics scans using machine learning. METHODS: We acquired 33 SD PET images, of which 13 had actual LD (ALD) PET, and simulated LD (SLD) PET images at seven different count levels from the SD PET scans. We employed image quality transfer (IQT), a machine learning algorithm that performs patch-regression to map parameters from low-quality to high-quality images. At each count level, patches extracted from 23 pairs of SD/SLD PET images were used to train three IQT models - global linear, single tree, and random forest regressions with cubic patch sizes of 3 and 5 voxels. The models were then used to estimate SD images from LD images at each count level for 10 unseen subjects. Lesion-detection task was carried out on matched lesion-present and lesion-absent images. RESULTS: LD PET-CT protocol yielded lesion detectability with sensitivity of 0.98 and specificity of 1. Random forest algorithm with cubic patch size of 5 allowed further 11.7% reduction in the effective PETCT dose without compromising lesion detectability, but underestimated SUV by 30%. CONCLUSION: LD PET-CT protocol was validated for lesion detection using ALD PET scans. Substantial image quality improvement or additional dose reduction while preserving clinical values can be achieved using machine learning methods though SUV quantification may be biased and adjustment of our research protocol is required for clinical use.

Evaluation of Multimodal Algorithms for the Segmentation of Multiparametric MRI Prostate Images.

Prostate segmentation in multiparametric magnetic resonance imaging (mpMRI) can help to support prostate cancer diagnosis and therapy treatment. However, manual segmentation of the prostate is subjective and time-consuming. Many deep learning monomodal networks have been developed for automatic whole prostate segmentation from T2-weighted MR images. We aimed to investigate the added value of multimodal networks in segmenting the prostate into the peripheral zone (PZ) and central gland (CG). We optimized and evaluated monomodal DenseVNet, multimodal ScaleNet, and monomodal and multimodal HighRes3DNet, which yielded dice score coefficients (DSC) of 0.875, 0.848, 0.858, and 0.890 in WG, respectively. Multimodal HighRes3DNet and ScaleNet yielded higher DSC with statistical differences in PZ and CG only compared to monomodal DenseVNet, indicating that multimodal networks added value by generating better segmentation between PZ and CG regions but did not improve the WG segmentation. No significant difference was observed in the apex and base of WG segmentation between monomodal and multimodal networks, indicating that the segmentations at the apex and base were more affected by the general network architecture. The number of training data was also varied for DenseVNet and HighRes3DNet, from 20 to 120 in steps of 20. DenseVNet was able to yield DSC of higher than 0.65 even for special cases, such as TURP or abnormal prostate, whereas HighRes3DNet's performance fluctuated with no trend despite being the best network overall. Multimodal networks did not add value in segmenting special cases but generally reduced variations in segmentation compared to the same matched monomodal network.

Convolutional neural networks for improving image quality with noisy PET data.

GOAL: PET is a relatively noisy process compared to other imaging modalities, and sparsity of acquisition data leads to noise in the images. Recent work has focused on machine learning techniques to improve PET images, and this study investigates a deep learning approach to improve the quality of reconstructed image volumes through denoising by a 3D convolution neural network. Potential improvements were evaluated within a clinical context by physician performance in a reading task. METHODS: A wide range of controlled noise levels was emulated from a set of chest PET data in patients with lung cancer, and a convolutional neural network was trained to denoise the reconstructed images using the full-count reconstructions as the ground truth. The benefits, over conventional Gaussian smoothing, were quantified across all noise levels by observer performance in an image ranking and lesion detection task. RESULTS: The CNN-denoised images were generally ranked by the physicians equal to or better than the Gaussian-smoothed images for all count levels, with the largest effects observed in the lowest-count image sets. For the CNN-denoised images, overall lesion contrast recovery was 60% and 90% at the 1 and 20 million count levels, respectively. Notwithstanding the reduced lesion contrast recovery in noisy data, the CNN-denoised images also yielded better lesion detectability in low count levels. For example, at 1 million true counts, the average true positive detection rate was around 40% for the CNN-denoised images and 30% for the smoothed images. CONCLUSION: Significant improvements were found for CNN-denoising for very noisy images, and to some degree for all noise levels. The technique presented here offered however limited benefit for detection performance for images at the count levels routinely encountered in the clinic.

18F-DCFPyL PET/CT in Patients with Subclinical Recurrence of Prostate Cancer: Effect of Lesion Size, Smoothing Filter, and Partial-Volume Correction on PROMISE Criteria.

Our purpose was to determine the effect of a smoothing filter and partial-volume correction (PVC) on measured prostate-specific membrane antigen (PSMA) activity in small metastatic lesions and to determine the impact of these changes on molecular imaging PSMA (miPSMA) scoring. Methods: Men who had biochemical recurrence of prostate cancer with negative findings on CT and bone scintigraphy were referred for 18F-DCFPyL (2-(3-(1-carboxy-5-[(6-18F-fluoro-pyridine-3-carbonyl)-amino]-pentyl) PET/CT. Examinations were performed on 1 of 2 different brands of PET/CT scanner. All suspected tumor sites were manually contoured on coregistered CT and PET images, and each was assigned an miPSMA score as per the PROMISE criteria. The PVC factors were calculated for every lesion using the anatomic CT and then applied to the unsmoothed PET images. The miPSMA scores, with and without the corrections, were compared, and a simplified rule-of-thumb (RoT) correction factor (CF) was derived for lesions at various sizes (<4 mm, 4-7 mm, 7-9 mm, and 9-12 mm). This CF was then applied to the original dataset and the miPSMA scores that were obtained using the RoT CF were compared with those obtained using the actual corrections. Results: There were 75 men (median age, 69 y; median serum PSA, 3.69 µg/L) with 232 metastatic nodes less than 12 mm in diameter (mean lesion volume, 313.5 ± 309.6 mm3). The mean SUVmax before and after correction was 11.0 ± 9.3 and 28.5 ± 22.8, respectively (P < 0.00001). The mean CF for lesions smaller than 4 mm (n = 22), 4-7 mm (n = 140), 7-9 mm (n = 50), and 9-12 mm (n = 20) was 4 (range, 2.5-6.4), 2.8 (range, 1.6-4.9), 2.3 (range, 1.6-3.3), and 1.8 (range, 1.4-2.4), respectively. Overall, the miPSMA scores were concordant between the corrected dataset and the RoT dataset for 205 of 232 lesions (88.4%). Conclusion: A smoothing filter and PVC had a significant effect on measured PSMA activity in small nodal metastases, impacting the miPSMA score.

Low dose positron emission tomography emulation from decimated high statistics: A clinical validation study.

PURPOSE: The fundamental nature of positron emission tomography (PET), as an event detection system, provides some flexibility for data handling, including retrospective data manipulation. The reorganization of acquisition data allows the emulation of new scans arising from identical radiotracer spatial distributions, but with different statistical compositions, and is especially useful for evaluating the stability and reproducibility of reconstruction algorithms or when investigating extremely low count conditions. This approach is ubiquitous in the research literature but has only been validated, from the point of view of the noise properties, with numerical simulations and phantom data. We present here the first experiment comparing PET images of the same human subjects generated with two separate injections of radiotracer, using actual low dose (LD) data to validate a randomly decimated emulation from a standard dose scan. A key point of the work is focused on the randoms fractions, which scale differently than the trues at varying activity levels. METHODS: Eleven patients with non-small cell lung cancer were enrolled in the study. Each imaging session consisted of two independent FDG-PET/CT scans: a LD scan followed by a standard dose (SD) scan. Images were first reconstructed, using filtered back-projection (FBP) and OSEM incorporating time-of-flight information and point-spread function modeling (PSFTOF), from the LD and SD datasets comprising all counts from each scanned bed position. The number of true counts was recorded for all LD scans, and independent, count-matched emulations (ELD) were reconstructed from the SD data. Noise distribution within the liver and standardized uptake value reproducibility within a population of contoured, tracer-avid lesion volumes were evaluated across scans and statistics. RESULTS: The randoms fraction estimates were 17.4 ± 1.6% (14.9-19.4) in the LD data and 42 ± 2.3% (37.1-45.5) in the SD data. Eleven lesions were identified and volumes of interest were generated with a 50% threshold isocontour for each lesion, in every image. The distributions of metabolic volumes, means and maxima defined by the contoured volumes-of-interest (VOIs) were similar between the LD and SD sets. A two-tailed, matched t-test was performed on the populations of region statistics for both LD and ELD reconstructions, and the t-statistics were 1.1 (P = 0.267) and -0.22 (P = 0.828) for the background liver VOIs and -0.54 (P = 0.603) and 0.23 (P = 0.821) for the lesion VOIs, for FBP and PSFTOF respectively. In every test, the null hypothesis that the two populations had the same mean could not be rejected at the 5% significance level. CONCLUSIONS: Our results demonstrate that clinical LD PET scans can indeed be accurately emulated by the statistical decimation of standard dose scans, and this was achieved through validation by images generated with unbiased random coincidence estimations.

The use of multiple time point dynamic positron emission tomography/computed tomography in patients with oral/head and neck cancer does not predictably identify metastatic cervical lymph nodes.

PURPOSE: To determine whether the time course of 18-fluorine fluorodeoxyglucose (18F-FDG) activity in multiple consecutively obtained 18F-FDG positron emission tomography (PET)/computed tomography (CT) scans predictably identifies metastatic cervical adenopathy in patients with oral/head and neck cancer. It is hypothesized that the activity will increase significantly over time only in those lymph nodes harboring metastatic cancer. PATIENTS AND METHODS: A prospective cohort study was performed whereby patients with oral/head and neck cancer underwent consecutive imaging at 9 time points with PET/CT from 60 to 115 minutes after injection with (18)F-FDG. The primary predictor variable was the status of the lymph nodes based on dynamic PET/CT imaging. Metastatic lymph nodes were defined as those that showed an increase greater than or equal to 10% over the baseline standard uptake values. The primary outcome variable was the pathologic status of the lymph node. RESULTS: A total of 2,237 lymph nodes were evaluated histopathologically in the 83 neck dissections that were performed in 74 patients. A total of 119 lymph nodes were noted to have hypermetabolic activity on the 90-minute (static) portion of the study and were able to be assessed by time points. When we compared the PET/CT time point (dynamic) data with the histopathologic analysis of the lymph nodes, the sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 60.3%, 70.5%, 66.0%, 65.2%, and 65.5%, respectively. CONCLUSIONS: The use of dynamic PET/CT imaging does not permit the ablative surgeon to depend only on the results of the PET/CT study to determine which patients will benefit from neck dissection. As such, we maintain that surgeons should continue to rely on clinical judgment and maintain a low threshold for executing neck dissection in patients with oral/head and neck cancer, including those patients with N0 neck designations.