Go-previous
10:30   Oncology - II
Automatic breast tumor identification in ultrafast MRI using temporal and spatial information
Xueping Jing, Mirjam Wielema, Paul E. Sijens, Matthijs Oudkerk, Monique Dorrius, Peter van Ooijen
Abstract: Purpose: The diagnosis of breast cancer with MRI is based on both morphological evaluation and kinetic curve assessment. Current computer-aided diagnosis methods for malignancy determination mainly focus on morphology features but ignore the temporal information in dynamic contrast-enhanced MRI sequences. The aim of this study is to investigate the feasibility of using deep learning methods to differentiate benign from malignant lesions in ultrafast MRI with both temporal and spatial information. Materials and Methods: 173 single breasts with suspicious findings above 5mm were retrospectively selected from 151 exams in 122 patients. Of those breasts, 109 contained benign lesions, and 64 contained malignant lesions. MIP images were generated from each sequence in the ultrafast MRI scan. A 2D convolutional neural network (CNN) and long short-term memory (LSTM) network were employed to learn morphological and temporal features, respectively. The 2D CNN model was trained with the MIPs from the last 4 scans to ensure the visibility of the lesion, while the LSTM model took MIPs of an entire scan as input. The 2D CNN and LSTM models were integrated by fusion of the predicted risk score from each model. The performance of each model and their combination was evaluated with a 100-times repeated stratified 4-fold cross-validation. Results: The 2D CNN showed a mean AUC of 0.81±0.06, the LSTM network showed a mean AUC of 0.78±0.07, while the combination of both showed a mean AUC of 0.83±0.06 in the repeated stratified cross-validation. Conclusion: Deep learning-based method using features derived from ultrafast MRI could help distinguish benign and malignant lesions. 2D CNN achieved a higher mean AUC than LSTM in the classification of benign and malignant lesions by ultrafast MRI, but the combined prediction could further improve the performance. There is clearly an added value in using the temporal information extracted by the LSTM model.
Parametric imaging of PSMA-targeted nanobubbles by improved intravascular modeling
Chuan Chen
Abstract: Introduction With a diameter 10 times smaller than microbubbles (MBs), nanobubbles (NBs) are novel ultrasound contrast agents that can permeate the vessel wall and reach specific receptors on the cell membrane [1, 2]. NBs targeted to the prostate-specific membrane antigen (PSMA), a well-studied biomarker overexpressed in prostate cancer (PCa), can actively bind to the receptors on PCa cells. The specific binding of PSMA-targeted NBs (PSMA-NBs) was demonstrated by a prolonged retention effect (PRE) in dual-tumor mice models [2]. In relevant works, the PRE was mainly analyzed by qualitatively or semi-quantitatively interpreting time-intensity curves (TICs) in a ROI. Although the pixel-level pharmacokinetics modeling of PSMA- NBs was shown to be feasible in our previous work [3], this work aims to further improve the pixel-level analysis by dedicated modeling of the NBs intravascular (plasma) input. Method Six mice were implanted with a PC3pip (PSMA+) and a PC3flu (PSMA-) tumor in the two contralateral flanks. After injecting a 200-μL bolus of untargeted NBs or PSMA-NBs, a PLT-1204BT probe (frequency, 12 MHz; MI, 0.1; Toshiba) was fixed to image the two tumors at 0.2 fps for 30 mins. TICs from each pixel in the delineated tumor ROIs were preprocessed by spatial filtering (Gaussian kernel size 0.5 mm). We adopted a general model for TIC fitting: I(t)=I_p (t)+βI_p (t)*exp⁡(-λt), where I_p (t) represents the plasma input, and β and λ denote the amplitude and decay rate of retention transfer, respectively [4]. For I_p (t), we compared two models: a) the modified local density random walk model (mLDRW-input) [5] known for accurate modeling of MBs flow and b) the gamma-variate model (gamma-input) [3]. Parametric images were estimated by non-linear least-squares fitting of pixel-level TICs by the two models, as illustrated in Fig. 1(a). Results and Discussion In Fig. 1(b), the estimated β and β/λ by the two models collected from six mice are compared by boxplots and Cohen's d [6] effect size. When employing the mLDRW-input model, both parameters show more significant differences between PSMA- and PSMA+ ROIs for PSMA-NBs, and between PSMA-NBs and untargeted NBs for PSMA+ ROIs. Therefore, this work suggests mLDRW modeling of the plasma input can produce better differentiation between different ROIs and NB types with less variance, in comparison to gamma-variate modeling. It can better interpreting of the PRE caused by extravascular binding at the pixel level.
Rigid and flexible multibody modelling of flexible instruments used in cervical cancer brachytherapy
Robin Straathof, Jaap Meijaard, Sharline van Vliet-Pérez, Inger-Karine Kolkman-Deurloo, Remi Nout, Ben Heijmen, Linda Wauben, Jenny Dankelman, Nick van de Berg
Abstract: Purpose In brachytherapy (BT) for cervical cancer, a radioactive source is driven by cable through an intracavitary (IC) applicator placed in the vaginal and uterine cavity, and/or through interstitial (IS) catheters guided through the applicator and implanted in tissue. Accurate understanding of source cable and catheter behaviour is important due to the steep dose gradient in BT of up to 12% per mm. The purpose of this study is to develop and validate comprehensible computer models to simulate: (1) BT source paths, and (2) insertion forces of catheters in curved IC/IS applicator channels. These models can aid novel (3D-printed) BT applicator development and improve source path models used for treatment planning. Method Source cables or catheters were modelled as an interconnected series of flexible beam elements or rigid beam elements connected through torsional springs. For evaluating the source cable model, simulated paths of a Flexitron source cable (Elekta, Stockholm, Sweden) in CT/MR ring applicators (Elekta, diameters: Ø26, Ø30 and Ø34 mm, angles: 45° and 60°) were compared with centreline data and manufacturer-specified source paths. For validating catheter models, simulated ProGuide 6F catheter with obturator (Elekta) insertion forces in S-shaped channels with varying design parameters (curvature, geometric torsion, and clearance) were compared with force measurements in dedicated 3D-printed templates. Results and discussion Rigid and flexible multibody models produced similar results. Maximum differences between dwell positions of the simulated source path and centreline varied between 4.0-6.4 mm in ring applicators of different sizes and were observed at the most distal dwell position. Simulated paths were in closer agreement with manufacturer-specified paths, with maximum differences of 0.7-1.4 mm between distal dwell positions. Insertion force simulation results of catheters were in close agreement with experimental results for all design parameters, and predicted peak forces were within 25%. The accuracy of simulated forces can improve by experimental determination and incorporation of friction coefficients. Conclusion The developed models show promising results in predicting the behaviour of flexible instruments in BT applicators. Insights from these models can aid novel applicator design with improved motion and force transmission of BT instruments, and contribute to overall treatment precision.
Acoustic characterization of tissue-mimicking materials for ultrasound perfusion imaging
Peiran Chen, Andreas Pollet, Anastasiia Panfilova, Meiyi Zhou, Simona Turco, Jaap den Toonder, Massimo Mischi
Abstract: Materials with well-characterized acoustic properties are of great interest for the development of tissue-mimicking phantoms with designed (micro)vasculature networks. These are important for experimental studies on ultrasound. Typical (micro)vasculature phantom materials are soft hydrogels that have low mechanical stiffness. To overcome this issue, the phantoms can be enclosed in a rigid case, avoiding deformation of the inside soft hydrogel while being imaged. This is of particular interest for research in ultrasound perfusion imaging as it prevents changes of position and geometry of the (micro)vasculature while imaging, and ensures reliable perfusion in the phantoms. To this end, we characterized the acoustic properties of a series of phantom and case materials, and provided considerations and recommendations for the choice of the optimal materials [1]. Based on these materials, perfusable phantoms can be fabricated by wire casting or sugar printing [2]. The through-transmission technique was used to characterize the acoustic properties of the phantom and case materials, using water as a reference. The transmitted frequency ranged from 1 MHz to 6 MHz for measuring the speed of sound (SoS), acoustic impedance (Z), and attenuation coefficient (α). The nonlinearity was assessed by determining the nonlinearity parameter B/A, using a transmit frequency of 2.25 MHz. The candidate materials were seven phantom materials at different compositions and five case materials. Perfusable phantoms and their ultrasound perfusion imaging were also demonstrated. Among these phantom materials, polyacrylamide (PAA) has the measured SoS ranging from 1445 to 1538 m/s, Z ranging from 1.45 to 1.58 *10^6 kg/m^2s, α ranging from about 0.1 to 0.9 dB/cm at frequencies varying from 1 MHz to 6 MHz, and B/A ranging from 6.1 to 11.6, which are close to the acoustic properties of human soft tissues. Moreover, PAA has good optical transparency, longevity and stability. Among these case materials, polymethylpentene (TPX) has the measured SoS and Z similar to that of soft tissues. Especially, its low α ranging from about 3.3 to 11.3 dB/cm would be the preferred choice. Therefore, our measurements suggest PAA and TPX to be the optimal phantom and case materials, respectively. [1] Chen et al., Ultrasound Med. Biol., 2021. [2] Pollet et al., Micromachines, 2019.
Machine learning models to predict risk on cancer-related fatigue in breast cancer patients
Lian Beenhakker, Kim Wijlens, Annemieke Witteveen, Marianne Heins, Christina Bode, Sabine Siesling, Miriam Vollenbroek-Hutten
Abstract: Background The survival rate after breast cancer has been increasing over the years, resulting in more survivors and more survivors experiencing long-term effects of cancer and its treatment. One of these is cancer-related fatigue (CRF), which is still experienced by 30% of the patients ten years after diagnosis. To enable early treatment of CRF to prevent the fatigue from becoming chronic, high-risk individuals should be identified and possibly monitored actively. The goal of this study is to explore possibilities to predict the risk an individual patient has for developing CRF. Methods Two datasets were obtained through the Primary Secondary Cancer Care Registry (PSCCR) with data on patient, tumour, treatment characteristics and late effects. For the first, PSCCR-PROFIEL, 392 patients and 24 input variables were used to predict self-reported fatigue yes/no. For the second, PSCCR, 12.813 patients and 64 input variables, including GP information of before breast cancer diagnosis, were used to predict fatigue complaints as registered by GPs. Missing data was imputed using Multiple Imputed Chained Equations. Several machine learning models were used: Random Forest Classifier, Logistic Regression, Gaussian Naïve Bayes, K-Nearest Neighbours and Multi-Layer Perceptron. For comparison, a statistical logistic regression model was developed. Models were optimized using a cross-fold grid search and output probabilities were compared using the area under the receiver operator characteristic curve (AUC-score). Results For the PSCCR-PROFIEL dataset, the best AUC-score was 0.67±0.04 using the logistic regression machine learning model. For the PSCCR dataset, the best AUC-score was 0.63±0.02 using the Random Forest Classifier. The statistical logistic regression model was best for PSCCR-PROFIEL with an AUC-score of 0.67±0.03. Visual comparison of the predicted probabilities did not show difference between fatigued and non-fatigued patients. Conclusion The current models created for the prediction of CRF do not show high discriminative ability. More differentiation in the outcome variable fatigue (such as Likert-scales or different fatigue domains) could lead to improved outcomes. Also, data on additional predictors could be needed to predict CRF. In a future study, we will collect more data to hopefully improve the predictions for CRF.
In vitro visualization of hepatic blood flow during radioembolization: does the catheter influence local hemodynamics?
Hadi Mirgolbabaee, Camille Kuitert, Kartik Jain, Guillaume Lajoinie, Frank Nijsen, Michel Versluis, Erik Groot Jebbink
Abstract: Liver tumors are increasingly treated using radioembolization (RE) treatment, in which radioactive microspheres are injected through a catheter placed in the hepatic arterial vasculature with the aim of embolizing and irradiating liver malignancies [1]. Although blood flow influences the distribution of microspheres inside the liver vasculature, it is not the governing parameter controlling the RE procedure [2]. Recent studies depicted that the microsphere distribution is influenced by catheter type and its position [1,3]. It is, however, unknown to what extent the catheter itself induces flow disturbances potentially leading to differences in microsphere distribution. Therefore, the goal is to study the effect of a catheter on the local blood flow dynamics. Two-dimensional (2D) laser particle image velocimetry (PIV) was used to quantify the flow field inside a simplified right hepatic artery (RHA) phantom. Due to the out-of-plane movement of the commercial catheter within the lumen, a stainless steel capillary pipe was fixed in the center of the lumen to represent the catheter. Initially, the flow field in the native lumen without catheter was quantified. Next, the catheter was placed in two locations; upstream in the RHA and near the apex of the bifurcation. PIV measurements were performed with and without injection through the catheter. This study revealed that the addition of a catheter in the flow lumen drastically influences the local velocity fields in the vicinity of the catheter’s tip. Moreover, the skewed radial velocity profiles in daughter branches were altered significantly by the jet stream caused by the catheter injection, specifically when the catheter was placed near the apex of the bifurcation. Furthermore, the emanated jet flow from the catheter showed an oscillatory behavior caused by the main flow pulsatility and the catheter position. Thus, the catheter location with respect to the bifurcation and its tip direction severely impact the local flow fields. The obtained results from this study will be used to validate our ongoing computational fluid dynamics (CFD) studies, in which more catheter positions will be simulated. These simulations can be employed in optimizing the RE treatment by pre-operative prediction of microsphere distribution in patient-specific liver vasculatures.


end %-->