Quantitative ultrasound spectroscopic imaging for characterization of disease extent in prostate cancer patients

Multiparametric Ultrasound Imaging of Prostate Cancer Using Deep Neural Networks

OBJECTIVE

A deep neural network (DNN) was trained to generate a multiparametric ultrasound (mpUS) volume from four input ultrasound-based modalities (acoustic radiation force impulse [ARFI] imaging, shear wave elasticity imaging [SWEI], quantitative ultrasound-midband fit [QUS-MF], and B-mode) for the detection of prostate cancer.

METHODS

A DNN was trained using co-registered ARFI, SWEI, MF, and B-mode data obtained in men with biopsy-confirmed prostate cancer prior to radical prostatectomy (15 subjects, comprising 980,620 voxels). Data were obtained using a commercial scanner that was modified to allow user control of the acoustic beam sequences and provide access to the raw image data. For each subject, the index lesion and a non-cancerous region were manually segmented using visual confirmation based on whole-mount histopathology data.

RESULTS

In a prostate phantom, the DNN increased lesion contrast-to-noise ratio (CNR) compared to a previous approach that used a linear support vector machine (SVM). In the in vivo test datasets (n = 15), the DNN-based mpUS volumes clearly portrayed histopathology-confirmed prostate cancer and significantly improved CNR compared to the linear SVM (2.79 ± 0.88 vs. 1.98 ± 0.73, paired-sample t-test p < 0.001). In a sub-analysis in which the input modalities to the DNN were selectively omitted, the CNR decreased with fewer inputs; both stiffness- and echogenicity-based modalities were important contributors to the multiparametric model.

CONCLUSION

The findings from this study indicate that a DNN can be optimized to generate mpUS prostate volumes with high CNR from ARFI, SWEI, MF, and B-mode and that this approach outperforms a linear SVM approach.

Evaluation of prostate cancer detection using micro-ultrasound versus MRI through co-registration to whole-mount pathology

Micro-ultrasound has recently been introduced as a low-cost alternative to multi-parametric MRI for imaging prostate cancer. Early clinical studies have demonstrated promising results; however, robust validation via comparison with whole-mount pathology has yet to be achieved. Due to micro-ultrasound probe design and tissue deformation during scanning, it is difficult to accurately correlate micro-ultrasound imaging planes with ground truth whole-mount pathology slides. In this study, we developed a multi-step methodology to co-register micro-ultrasound and MRI to whole-mount pathology. The three-step process had a registration error of 3.90 ± 0.11 mm and consists of: (1) micro-ultrasound image reconstruction, (2) 3D landmark registration of micro-ultrasound to MRI, and (3) 2D capsule registration of MRI to whole-mount pathology. This process was then used in a preliminary reader study to compare the diagnostic accuracy of micro-ultrasound and MRI in 15 patients who underwent radical prostatectomy for prostate cancer. Micro-ultrasound was found to have equivalent performance to retrospective MRI review for index lesion detection (91.7% vs. 80%), while demonstrating an increased detection of tumor extent (52.5% vs. 36.7%) with similar false positive regions-of-interest (38.3% vs. 40.8%). Prospective MRI review had reduced detection of index lesions (73.3%) and tumor extent (18.9%) but improved false positive regions-of-interest (22.7%) relative to micro-ultrasound and retrospective MRI. Further evaluation is needed with a larger sample size.

Deep learning of quantitative ultrasound multi-parametric images at pre-treatment to predict breast cancer response to chemotherapy

In this study, a novel deep learning-based methodology was investigated to predict breast cancer response to neo-adjuvant chemotherapy (NAC) using the quantitative ultrasound (QUS) multi-parametric imaging at pre-treatment. QUS multi-parametric images of breast tumors were generated using the data acquired from 181 patients diagnosed with locally advanced breast cancer and planned for NAC followed by surgery. The ground truth response to NAC was identified for each patient after the surgery using the standard clinical and pathological criteria. Two deep convolutional neural network (DCNN) architectures including the residual network and residual attention network (RAN) were explored for extracting optimal feature maps from the parametric images, with a fully connected network for response prediction. In different experiments, the features maps were derived from the tumor core only, as well as the core and its margin. Evaluation results on an independent test set demonstrate that the developed model with the RAN architecture to extract feature maps from the expanded parametric images of the tumor core and margin had the best performance in response prediction with an accuracy of 88% and an area under the receiver operating characteristic curve of 0.86. Ten-year survival analyses indicate statistically significant differences between the survival of the responders and non-responders identified based on the model prediction at pre-treatment and the standard criteria at post-treatment. The results of this study demonstrate the promising capability of DCNNs with attention mechanisms in predicting breast cancer response to NAC prior to the start of treatment using QUS multi-parametric images.

Characterizing intra-tumor regions on quantitative ultrasound parametric images to predict breast cancer response to chemotherapy at pre-treatment

The efficacy of quantitative ultrasound (QUS) multi-parametric imaging in conjunction with unsupervised classification algorithms was investigated for the first time in characterizing intra-tumor regions to predict breast tumor response to chemotherapy before the start of treatment. QUS multi-parametric images of breast tumors were generated using the ultrasound radiofrequency data acquired from 181 patients diagnosed with locally advanced breast cancer and planned for neo-adjuvant chemotherapy followed by surgery. A hidden Markov random field (HMRF) expectation maximization (EM) algorithm was applied to identify distinct intra-tumor regions on QUS multi-parametric images. Several features were extracted from the segmented intra-tumor regions and tumor margin on different parametric images. A multi-step feature selection procedure was applied to construct a QUS biomarker consisting of four features for response prediction. Evaluation results on an independent test set indicated that the developed biomarker coupled with a decision tree model with adaptive boosting (AdaBoost) as the classifier could predict the treatment response of patient at pre-treatment with an accuracy of 85.4% and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.89. In comparison, the biomarkers consisted of the features derived from the entire tumor core (without consideration of the intra-tumor regions), and the entire tumor core and the tumor margin could predict the treatment response of patients with an accuracy of 74.5% and 76.4%, and an AUC of 0.79 and 0.76, respectively. Standard clinical features could predict the therapy response with an accuracy of 69.1% and an AUC of 0.6. Long-term survival analyses indicated that the patients predicted by the developed model as responders had a significantly better survival compared to the non-responders. Similar findings were observed for the two response cohorts identified at post-treatment based on standard clinical and pathological criteria. The results obtained in this study demonstrated the potential of QUS multi-parametric imaging integrated with unsupervised learning methods in identifying distinct intra-tumor regions in breast cancer to characterize its responsiveness to chemotherapy prior to the start of treatment.

A Brain Tumor Segmentation Framework Based on Outlier Detection Using One-Class Support Vector Machine

Accurate segmentation of brain tumors is a challenging task and also a crucial step in diagnosis and treatment planning for cancer patients. Magnetic resonance imaging (MRI) is the standard imaging modality for detection, characterization, treatment planning and outcome evaluation of brain tumors. MRI scans are usually acquired at multiple sessions before and after the treatment. An automatic segmentation framework is highly desirable to segment brain tumors in MR images as it streamlines the image-guided radiation therapy workflow considerably. Automatic segmentation of brain tumors also facilitates an incremental development of data-driven systems for therapy outcome prediction based on radiomics analysis. In this study, an outlier-detection-based segmentation framework is proposed to delineate brain tumors in magnetic resonance (MR) images automatically. The proposed method considers the tumor and edema pixels in an MR image as outliers compared to the pixels associated with the healthy tissue. The framework generates two outlier masks using independent one-class support vector machines that operate on post-contrast T1-weighted (T1w) and T2-weighted-fluid-attenuation-inversion-recovery (T2-FLAIR) images. The outlier masks are subsequently refined and fused using a number of morphological and logical operators to estimate a tumor mask for each image slice. The framework was constructed and evaluated using the MRI data acquired from 35 and 5 patients with brain metastasis, respectively. The obtained results demonstrated an average Dice similarity coefficient and Hausdorff distance of 0.84 ± 0.06 and 1.85 ± 0.48 mm, respectively, between the manual (ground truth) and automatic tumor contours, on the independent test set.

An Automatic Framework for Segmentation of Brain Tumours at Follow-up Scans after Radiation Therapy

Brain metastasis is the most common intracranial malignancy with a poor overall survival (OS) after treatment. The standard stereotactic radiation therapy (SRT) planning procedure for brain metastasis requires delineating the tumour volume on magnetic resonance (MR) images. MR images are also acquired at multiple follow-up scans after SRT to monitor the treatment outcome through measuring changes in the physical dimensions of the tumour. Such measurements require manual segmentation of the tumour volume on multiple slices of several follow-up images which is tedious and impedes the SRT evaluation work flow considerably. In this study, an automatic framework was proposed to segment the tumour volume on longitudinal MR images acquired at standard follow-up scans after SRT. The multi-step segmentation framework was based on region growing and morphological snakes models that applied the standard SRT planning tumour contour as a basis to approximate the tumour shape and location at each follow-up scan for an accurate automatic segmentation of tumour volume. The framework was evaluated using the MR imaging data acquired from five patients prior to and at three follow-up scans after SRT. The preliminary results indicated that the Dice similarity coefficient between the ground truth tumour masks and their automatically segmented counterparts ranged between 0.84 and 0.90, while the average Dice coefficient for all the follow-up scans was 0.88. The results obtained implied a good potential of the proposed framework for being incorporated into the SRT treatment planning and evaluation systems as well as outcome prediction models.

Quantitative MRI Biomarkers of Stereotactic Radiotherapy Outcome in Brain Metastasis

About 20-40% of cancer patients develop brain metastases, causing significant morbidity and mortality. Stereotactic radiation treatment is an established option that delivers high dose radiation to the target while sparing the surrounding normal tissue. However, up to 20% of metastatic brain tumours progress despite stereotactic treatment, and it can take months before it is evident on follow-up imaging. An early predictor of radiation therapy outcome in terms of tumour local failure (LF) is crucial, and can facilitate treatment adjustments or allow for early salvage treatment. In this study, an MR-based radiomics framework was proposed to derive and investigate quantitative MRI (qMRI) biomarkers for the outcome of LF in brain metastasis patients treated with hypo-fractionated stereotactic radiation therapy (SRT). The qMRI biomarkers were constructed through a multi-step feature extraction/reduction/selection framework using the conventional MR imaging data acquired from 100 patients (133 lesions), and were applied in conjunction with machine learning techniques for outcome prediction and risk assessment. The results indicated that the majority of the features in the optimal qMRI biomarkers characterize the heterogeneity in the surrounding regions of tumour including edema and tumour/lesion margins. The optimal qMRI biomarker consisted of five features that predict the outcome of LF with an area under the curve (AUC) of 0.79, and a cross-validated sensitivity and specificity of 81% and 79%, respectively. The Kaplan-Meier analyses showed a statistically significant difference in local control (p-value < 0.0001) and overall survival (p = 0.01). Findings from this study are a step towards using qMRI for early prediction of local failure in brain metastasis patients treated with SRT. This may facilitate early adjustments in treatment, such as surgical resection or salvage radiation, that can potentially improve treatment outcomes. Investigations on larger cohorts of patients are, however, required for further validation of the technique.

Quantitative Ultrasound Spectroscopy for Differentiation of Hepatocellular Carcinoma from At-Risk and Normal Liver Parenchyma

PURPOSE

Quantitative ultrasound approaches can capture tissue morphologic properties to augment clinical diagnostics. This study aims to clinically assess whether quantitative ultrasound spectroscopy (QUS) parameters measured in hepatocellular carcinoma (HCC) tissues can be differentiated from those measured in at-risk or healthy liver parenchyma.

EXPERIMENTAL DESIGN

This prospective Health Insurance Portability and Accountability Act (HIPAA)-compliant study was approved by the Institutional Review Board. Fifteen patients with HCC, 15 non-HCC patients with chronic liver disease, and 15 healthy volunteers were included (31.1% women; 68.9% men). Ultrasound radiofrequency data were acquired in each patient in both liver lobes at two focal depths (3/9 cm). Region of interests (ROIs) were drawn on HCC and liver parenchyma. The average normalized power spectrum for each ROI was extracted, and a linear regression was fit within the -6 dB bandwidth, from which the midband fit (MBF), spectral intercept (SI), and spectral slope (SS) were extracted. Differences in QUS parameters between the ROIs were tested by a mixed-effects regression.

RESULTS

There was a significant intraindividual difference in MBF, SS, and SI between HCC and adjacent liver parenchyma (P < 0.001), and a significant interindividual difference between HCC and at-risk and healthy non-HCC parenchyma (P < 0.001). In patients with HCC, cirrhosis (n = 13) did not significantly change any of the three parameters (P > 0.8) in differentiating HCC from non-HCC parenchyma. MBF (P = 0.12), SI (P = 0.33), and SS (P = 0.57) were not significantly different in non-HCC tissue among the groups.

CONCLUSIONS

The QUS parameters are significantly different in HCC versus non-HCC liver parenchyma, independent of underlying cirrhosis. This could be leveraged for improved HCC detection with ultrasound in the future.

Transbronchial evaluation of peripheral pulmonary lesions using ultrasonic spectrum analysis in lung cancer patients

BACKGROUND AND OBJECTIVE

Analysis of the endobronchial ultrasound (EBUS) radiofrequency spectrum has been used for convex-probe EBUS technology. Quantitative imaging analysis is also warranted for guided bronchoscopy using radial-probe EBUS (RP-EBUS) targeting peripheral pulmonary lesions (PPL). This study aimed to determine the feasibility of radiofrequency spectrum analysis for distinguishing malignant and benign PPL during diagnostic bronchoscopy.

METHODS

Raw RP-EBUS images with radiofrequency data, including backscatter signals, were prospectively recorded. The ultrasonic spectral parameters, such as intercept, midband-fit and slope within the region of interest, were retrospectively computed by linear regression analysis and compared with the final diagnosis.

RESULTS

A total of 71 PPL, including 45 malignant and 26 benign lesions, were analysed. Malignant PPL showed a significantly lower intercept (P < 0.0001), lower midband-fit (P < 0.0001) and higher slope (P = 0.014) than benign PPL. Analyses of the area under the curve of receiver operating characteristic plots demonstrated that the intercept showed the best diagnostic performance among three parameters (0.87, 0.77 and 0.69 for intercept, midband-fit and slope, respectively). The sensitivity, specificity, accuracy, positive likelihood and negative likelihood were 75.6%, 96.2%, 83.1%, 19.6 and 0.25 for the intercept; 88.9%, 57.7%, 77.5%, 2.1 and 0.19 for the midband-fit; and 68.9%, 73.1%, 70.4%, 2.6 and 0.43 for the slope.

CONCLUSION

Spectrum analysis of EBUS radiofrequency can be used as a novel non-invasive predictor of malignant or benign PPL. Analysis of the 'intercept' of the targeted lesion may provide useful supporting data for real-time sampling from PPL during diagnostic bronchoscopy.

Ultrasound Elastography of the Prostate Using an Unconstrained Modulus Reconstruction Technique: A Pilot Clinical Study

A novel full-inversion-based technique for quantitative ultrasound elastography was investigated in a pilot clinical study on five patients for non-invasive detection and localization of prostate cancer and quantification of its extent. Conventional-frequency ultrasound images and radiofrequency (RF) data (~5 MHz) were collected during mechanical stimulation of the prostate using a transrectal ultrasound probe. Pre and post-compression RF data were used to construct the strain images. The Young's modulus (YM) images were subsequently reconstructed using the derived strain images and the stress distribution estimated iteratively using finite element (FE) analysis. Tumor regions determined based on the reconstructed YM images were compared to whole-mount histopathology images of radical prostatectomy specimens. Results indicated that tumors were significantly stiffer than the surrounding tissue, demonstrating a relative YM of 2.5 ± 0.8 compared to normal prostate tissue. The YM images had a good agreement with the histopathology images in terms of tumor location within the prostate. On average, 76% ± 28% of tumor regions detected based on the proposed method were inside respective tumor areas identified in the histopathology images. Results of a linear regression analysis demonstrated a good correlation between the disease extents estimated using the reconstructed YM images and those determined from whole-mount histopathology images (r² = 0.71). This pilot study demonstrates that the proposed method has a good potential for detection, localization and quantification of prostate cancer. The method can potentially be used for prostate needle biopsy guidance with the aim of decreasing the number of needle biopsies. The proposed technique utilizes conventional ultrasound imaging system only while no additional hardware attachment is required for mechanical stimulation or data acquisition. Therefore, the technique may be regarded as a non-invasive, low cost and potentially widely-available clinical tool for prostate cancer diagnosis.

Photoacoustic signal characterization of cancer treatment response: Correlation with changes in tumor oxygenation

Frequency analysis of the photoacoustic radiofrequency signals and oxygen saturation estimates were used to monitor the in-vivo response of a novel, thermosensitive liposome treatment. The liposome encapsulated doxorubicin (HaT-DOX) releasing it rapidly (<20 s) when the tumor was exposed to mild hyperthermia (43 °C). Photoacoustic imaging (VevoLAZR, 750/850 nm, 40 MHz) of EMT-6 breast cancer tumors was performed 30 min pre- and post-treatment and up to 7 days post-treatment (at 2/5/24 h timepoints). HaT-DOX-treatment responders exhibited on average a 22% drop in oxygen saturation 2 h post-treatment and a decrease (45% at 750 nm and 73% at 850 nm) in the slope of the normalized PA frequency spectra. The spectral slope parameter correlated with treatment-induced hemorrhaging which increased the optical absorber effective size via interstitial red blood cell leakage. Combining frequency analysis and oxygen saturation estimates differentiated treatment responders from non-responders/control animals by probing the treatment-induced structural changes of blood vessel.

Interpretation of cardiac wall motion from cine-MRI combined with parametric imaging based on the Hilbert transform

OBJECT

The aim of this study was to test and validate the clinical impact of parametric amplitude images obtained using the Hilbert transform on the regional interpretation of cardiac wall motion abnormalities from cine-MR images by non-expert radiologists compared with expert consensus.

MATERIALS AND METHODS

Cine-MRI short-axis images obtained in 20 patients (10 with myocardial infarction, 5 with myocarditis and 5 with normal function) were processed to compute a parametric amplitude image for each using the Hilbert transform. Two expert radiologists blindly reviewed the cine-MR images to define a gold standard for wall motion interpretation for each left ventricular sector. Two non-expert radiologists reviewed and graded the same images without and in combination with parametric images. Grades assigned to each segment in the two separate sessions were compared with the gold standard.

RESULTS

According to expert interpretation, 264/320 (82.5%) segments were classified as normal and 56/320 (17.5%) were considered abnormal. The accuracy of the non-expert radiologists' grades compared to the gold standard was significantly improved by adding parametric images (from 87.2 to 94.6%) together with sensitivity (from 64.29 to 84.4%) and specificity (from 92 to 96.9%), also resulting in reduced interobserver variability (from 12.8 to 5.6%).

CONCLUSION

The use of parametric amplitude images based on the Hilbert transform in conjunction with cine-MRI was shown to be a promising technique for improvement of the detection of left ventricular wall motion abnormalities in less expert radiologists.

Quantification of Ultrasonic Scattering Properties of In Vivo Tumor Cell Death in Mouse Models of Breast Cancer

INTRODUCTION: Quantitative ultrasound parameters based on form factor models were investigated as potential biomarkers of cell death in breast tumor (MDA-231) xenografts treated with chemotherapy. METHODS: Ultrasound backscatter radiofrequency data were acquired from MDA-231 breast cancer tumor–bearing mice (n = 20) before and after the administration of chemotherapy drugs at two ultrasound frequencies: 7 MHz and 20 MHz. Radiofrequency spectral analysis involved estimating the backscatter coefficient from regions of interest in the center of the tumor, to which form factor models were fitted, resulting in estimates of average scatterer diameter and average acoustic concentration (AAC). RESULTS: The ∆AAC parameter extracted from the spherical Gaussian model was found to be the most effective cell death biomarker (at the lower frequency range, r² = 0.40). At both frequencies, AAC in the treated tumors increased significantly (P = .026 and .035 at low and high frequencies, respectively) 24 hours after treatment compared with control tumors. Furthermore, stepwise multiple linear regression analysis of the low-frequency data revealed that a multiparameter quantitative ultrasound model was strongly correlated to cell death determined histologically posttreatment (r² = 0.74). CONCLUSION: The Gaussian form factor model–based scattering parameters can potentially be used to track the extent of cell death at clinically relevant frequencies (7 MHz). The 20-MHz results agreed with previous findings in which parameters related to the backscatter intensity (i.e., AAC) increased with cell death. The findings suggested that, in addition to the backscatter coefficient parameter ∆AAC, biological features including tumor heterogeneity and initial tumor volume were important factors in the prediction of cell death response.