Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

Cardiology Cardiology

Impact of Donor Hemodynamics on Heart Transplant Outcomes: Using Machine Learning Techniques.

In The Journal of heart and lung transplantation : the official publication of the International Society for Heart Transplantation

PURPOSE : Several studies have evaluated the role of recipient hemodynamics on heart transplant (HTx) outcomes. Data on the role of donor hemodynamics and donor-related characteristics on outcomes is scarce.

METHODS : We studied adult (≥18 years) Htx patients from 1997 and 2016 using the UNOS database that had donor right heart catheterization information. Primary study endpoint was the composite of 1-year mortality or re-transplant. We utilized Machine Learning (ML) method and Logistic Regression (LR) to explore predictive variables associated with the primary outcome. We excluded post-transplant variables, donors with missing demographic data and improbable hemodynamics. Receiver operating characteristic (ROC) curve was constructed to investigate the discrimination ability of the model.

RESULTS : We included 4,053 Htx recipients where donors had received right heart catheterization. Median recipient age was 55 years and 76.7% were male. Prominent donor characteristics included a median donor age of 36 years (IQR 25-45 years), 69.3% male sex, 69.5% Caucasian by race, 22.8% with smoking history, 15.1% with heavy alcohol use, and 17.3% had known cocaine use. The ML models reasonably predicted one-year mortality or retransplantation (XGBoost AUC 0.64, AutoML AUC 0.66). The LR model showed a poorer predictability of the primary study endpoint (AUC 0.60). Recipient characteristics such as total bilirubin, serum creatinine, ventilator use, and donor characteristics such as donor age and right ventricular mass were predictive of outcomes in both models. No additional donor characteristics including hemodynamics predicted the primary endpoint.

CONCLUSION : Donor age and donor right ventricular mass were predictive of outcomes in both ML and LR models. Other donor characteristics including hemodynamics did not improve model predictability. Recipient characteristics significantly outweigh those of donor in predicting one-year mortality or retransplantation following Htx.

Shah M, Villela M A, Bravo C, Castellanos A


Radiology Radiology

The delineation of largely deformed brain midline using regression-based line detection network.

In Medical physics ; h5-index 59.0

PURPOSE : The human brain has two cerebral hemispheres that are roughly symmetric and separated by a midline, which is nearly a straight line shown in axial CT images in healthy subjects. However, brain diseases such as hematoma and tumors often cause midline shift, where the degree of shift can be regarded as a quantitative indication in clinical practice. To facilitate clinical evaluation, we need computer-aided methods to automate this quantification. Nevertheless, most existing studies focused on the landmark- or symmetry-based methods that provide only the existence of shift or its maximum distance, which could be easily affected by anatomical variability and large brain deformations. Intuitive results such as midline delineation or measurement are lacking. In this study, we focus on developing an automated and robust method based on the fully convolutional neural network for the delineation of midline in largely deformed brains.

METHODS : We propose a novel regression-based line detection network (RLDN) for the robust midline delineation, especially in largely deformed brains. Specifically, to improve the robustness of delineation in largely deformed brains, we regard the delineation of the midline as the skeleton extraction task and then use the multi-scale bidirectional integration module to acquire more representative features. Based on the skeleton extraction, we incorporate the regression task into it to delineate more accurate and continuous midline, especially in largely deformed brains. Our study utilized the public CQ 500 dataset (128 subjects) for training with hold-out validation on 61 subjects from a private cohort accrued from a local hospital.

RESULTS : The mean line distance error and F1-score were 1.17 ± 0.72 mm with 0.78 on CQ 500 test set, and 4.15 ± 3.97 mm with 0.61 on the private dataset. Besides, significant differences (p < 0.05) were observed between our method and other comparative ones on these two datasets.

CONCLUSIONS : This work provides a novel solution to acquire robust delineation of the midline, especially in largely deformed brains, and achieves state-of-the-art performance on the public and our private dataset, which makes it possible for automated diagnosis of relevant brain diseases in the future.

Wei Hao, Tang Xiangyu, Zhang Minqing, Li Qingfeng, Xing Xiaodan, Zhou Xiang Sean, Xue Zhong, Zhu Wenzhen, Chen Zailiang, Shi Feng


automated midline delineation, brain CT, brain midline shift, deep learning

Surgery Surgery

Prediction of clinical height gain from surgical posterior correction of idiopathic scoliosis.

In Journal of neurosurgery. Spine

OBJECTIVE : The best predictors of height gain due to surgical correction are the number of fused vertebrae and the degrees of the corrected Cobb angle. Existing studies of predictive models measured the radiographic spinal height and did not report the clinical height gain. The aims of this study were to determine the best predictive factors of clinical height gain before surgical correction, construct a predictive model using patient population data for machine learning, and test the performance of this model on a validation population.

METHODS : The authors reviewed 145 medical records of consecutive patients who underwent surgery that included placement of posterior spinal instrumentation and fusion for idiopathic scoliosis between 2012 and 2016. Standing and sitting clinical heights were measured before and after surgery in patients who had been surgically treated under similar conditions. Multivariate analysis was then performed and the results were used to develop a predictive model for height gain after surgery. The data from the included patients were randomly assigned to a learning set or a test set.

RESULTS : In total, 116 patients were included in the analysis, for whom the average postoperative clinical height gain in a standing position was 4.2 ± 1.8 cm (range 0-11 cm). The best prediction model was calculated as follows: standing clinical height gain (cm) = 1 - 0.023 × sitting clinical height (cm) - 0.19 × Risser stage + 0.058 × Cobb preoperative angle (°) + 0.021 × T5-12 kyphosis (°) + 0.14 × number of levels fused. In the validation cohort, 91% of the predicted values had an error of less than one-half of the actual height gain.

CONCLUSIONS : This predictive model formula for calculating the potential postoperative height gain after surgical treatment can be used preoperatively to inform idiopathic scoliosis patients of what outcomes they may expect from posterior spinal instrumentation and fusion (taking into account the model's uncertainty).

Langlais Tristan, Verdun Stephane, Compagnon Roxane, Ursu Catalin, Vergari Claudio, Barret Hugo, Morin Christian


adolescent, correction fusion, deformity, height, idiopathic, pediatrics, posterior instrumentation, prognosis study, scoliosis, spine

Surgery Surgery

Three-dimensional assessment of robot-assisted pedicle screw placement accuracy and instrumentation reliability based on a preplanned trajectory.

In Journal of neurosurgery. Spine

OBJECTIVE : Robotic spine surgery systems are increasingly used in the US market. As this technology gains traction, however, it is necessary to identify mechanisms that assess its effectiveness and allow for its continued improvement. One such mechanism is the development of a new 3D grading system that can serve as the foundation for error-based learning in robot systems. Herein the authors attempted 1) to define a system of providing accuracy data along all three pedicle screw placement axes, that is, cephalocaudal, mediolateral, and screw long axes; and 2) to use the grading system to evaluate the mean accuracy of thoracolumbar pedicle screws placed using a single commercially available robotic system.

METHODS : The authors retrospectively reviewed a prospectively maintained, IRB-approved database of patients at a single tertiary care center who had undergone instrumented fusion of the thoracic or lumbosacral spine using robotic assistance. Patients with preoperatively planned screw trajectories and postoperative CT studies were included in the final analysis. Screw accuracy was measured as the net deviation of the planned trajectory from the actual screw trajectory in the mediolateral, cephalocaudal, and screw long axes.

RESULTS : The authors identified 47 patients, 51% male, whose pedicles had been instrumented with a total of 254 screws (63 thoracic, 191 lumbosacral). The patients had a mean age of 61.1 years and a mean BMI of 30.0 kg/m2. The mean screw tip accuracies were 1.3 ± 1.3 mm, 1.2 ± 1.1 mm, and 2.6 ± 2.2 mm in the mediolateral, cephalocaudal, and screw long axes, respectively, for a net linear deviation of 3.6 ± 2.3 mm and net angular deviation of 3.6° ± 2.8°. According to the Gertzbein-Robbins grading system, 184 screws (72%) were classified as grade A and 70 screws (28%) as grade B. Placement of 100% of the screws was clinically acceptable.

CONCLUSIONS : The accuracy of the discussed robotic spine system is similar to that described for other surgical systems. Additionally, the authors outline a new method of grading screw placement accuracy that measures deviation in all three relevant axes. This grading system could provide the error signal necessary for unsupervised machine learning by robotic systems, which would in turn support continued improvement in instrumentation placement accuracy.

Jiang Bowen, Pennington Zach, Zhu Alex, Matsoukas Stavros, Ahmed A Karim, Ehresman Jeff, Mahapatra Smruti, Cottrill Ethan, Sheppell Hailey, Manbachi Amir, Crawford Neil, Theodore Nicholas


pedicle screw accuracy, robot assistance, surgical technique, three-dimensional accuracy

Radiology Radiology

Automated coronary artery atherosclerosis detection and weakly supervised localization on coronary CT angiography with a deep 3-dimensional convolutional neural network.

In Computerized medical imaging and graphics : the official journal of the Computerized Medical Imaging Society

We propose a fully automated algorithm based on a deep learning framework enabling screening of a coronary computed tomography angiography (CCTA) examination for confident detection of the presence or absence of coronary artery atherosclerosis. The system starts with extracting the coronary arteries and their branches from CCTA datasets and representing them with multi-planar reformatted volumes; pre-processing and augmentation techniques are then applied to increase the robustness and generalization ability of the system. A 3-dimensional convolutional neural network (3D-CNN) is utilized to model pathological changes (e.g., atherosclerotic plaques) in coronary vessels. The system learns the discriminatory features between vessels with and without atherosclerosis. The discriminative features at the final convolutional layer are visualized with a saliency map approach to provide visual clues related to atherosclerosis likelihood and location. We have evaluated the system on a reference dataset representing 247 patients with atherosclerosis and 246 patients free of atherosclerosis. With five fold cross-validation, an Accuracy = 90.9%, Positive Predictive Value = 58.8%, Sensitivity = 68.9%, Specificity of 93.6%, and Negative Predictive Value (NPV) = 96.1% are achieved at the artery/branch level with threshold 0.5. The average area under the receiver operating characteristic curve is 0.91. The system indicates a high NPV, which may be potentially useful for assisting interpreting physicians in excluding coronary atherosclerosis in patients with acute chest pain.

Candemir Sema, White Richard D, Demirer Mutlu, Gupta Vikash, Bigelow Matthew T, Prevedello Luciano M, Erdal Barbaros S


3D convolutional neural networks, Coronary artery computed tomography angiography, Coronary artery disease, Stenosis classification, Weakly supervised localization

General General

Progressive learning: A deep learning framework for continual learning.

In Neural networks : the official journal of the International Neural Network Society

Continual learning is the ability of a learning system to solve new tasks by utilizing previously acquired knowledge from learning and performing prior tasks without having significant adverse effects on the acquired prior knowledge. Continual learning is key to advancing machine learning and artificial intelligence. Progressive learning is a deep learning framework for continual learning that comprises three procedures: curriculum, progression, and pruning. The curriculum procedure is used to actively select a task to learn from a set of candidate tasks. The progression procedure is used to grow the capacity of the model by adding new parameters that leverage parameters learned in prior tasks, while learning from data available for the new task at hand, without being susceptible to catastrophic forgetting. The pruning procedure is used to counteract the growth in the number of parameters as further tasks are learned, as well as to mitigate negative forward transfer, in which prior knowledge unrelated to the task at hand may interfere and worsen performance. Progressive learning is evaluated on a number of supervised classification tasks in the image recognition and speech recognition domains to demonstrate its advantages compared with baseline methods. It is shown that, when tasks are related, progressive learning leads to faster learning that converges to better generalization performance using a smaller number of dedicated parameters.

Fayek Haytham M, Cavedon Lawrence, Wu Hong Ren


Computer vision, Continual learning, Deep learning, Machine learning, Neural networks, Speech recognition