Human Not in the Loop: Objective Sample Difficulty Measures for Curriculum Learning.

Curriculum learning is a learning method that trains models in a meaningful order from easier to harder samples. A key here is to devise automatic and objective difficulty measures of samples. In the medical domain, previous work applied domain knowledge from human experts to qualitatively assess classification difficulty of medical images to guide curriculum learning, which requires extra annotation efforts, relies on subjective human experience, and may introduce bias. In this work, we propose a new automated curriculum learning technique using the variance of gradients (VoG) to compute an objective difficulty measure of samples and evaluated its effects on elbow fracture classification from X-ray images. Specifically, we used VoG as a metric to rank each sample in terms of the classification difficulty, where high VoG scores indicate more difficult cases for classification, to guide the curriculum training process We compared the proposed technique to a baseline (without curriculum learning), a previous method that used human annotations on classification difficulty, and anti-curriculum learning. Our experiment results showed comparable and higher performance for the binary and multi-class bone fracture classification tasks.

Hanging protocol optimization of lumbar spine radiographs with machine learning.

OBJECTIVES: The purpose of this study was to determine whether machine learning algorithms can be utilized to optimize the hanging protocol of lumbar spine radiographs. Specifically, we explored whether machine learning models can accurately label lumbar spine views/positions, detect hardware, and rotate the lateral views to straighten the image. METHODS: We identified 1727 patients with 6988 lumbar spine radiographs. The view (anterior-posterior, right oblique, left oblique, left lateral, right lateral, left lumbosacral or right lumbosacral), hardware (present or not present), dynamic position (neutral, flexion, or extension), and correctional rotation of each radiograph were manually documented by a board-certified radiologist. Various output metrics were calculated, including area under the curve (AUC) for the categorical output models (view, hardware, and dynamic position). For non-binary categories, an all-versus-other technique was utilized designating one category as true and all others as false, allowing for a binary evaluation (e.g., AP vs. non-AP or extension vs. non-extension). For correctional rotation, the degree of rotation required to straighten the lateral spine radiograph was documented. The mean absolute difference was calculated between the ground truth and model-predicted value reported in degrees of rotation. Ensembles of the rotation models were created. We evaluated the rotation models on 3 test dataset splits: only 0 rotation, only non-0 rotation, and all cases. RESULTS: The AUC values for the categorical models ranged from 0.985 to 1.000. For the only 0 rotation data, the ensemble combining the absolute minimum value between the 20- and 60-degree models performed best (mean absolute difference of 0.610). For the non-0 rotation data, the ensemble merging the absolute maximum value between the 40- and 160-degree models performed best (mean absolute difference of 4.801). For the all cases split, the ensemble combining the minimum value of the 20- and 40-degree models performed best (mean absolute difference of 3.083). CONCLUSION: Machine learning techniques can be successfully implemented to optimize lumbar spine x-ray hanging protocols by accounting for views, hardware, dynamic position, and rotation correction.

Knowledge-Guided Multiview Deep Curriculum Learning for Elbow Fracture Classification.

Elbow fracture diagnosis often requires patients to take both frontal and lateral views of elbow X-ray radiographs. In this paper, we propose a multiview deep learning method for an elbow fracture subtype classification task. Our strategy leverages transfer learning by first training two single-view models, one for frontal view and the other for lateral view, and then transferring the weights to the corresponding layers in the proposed multiview network architecture. Meanwhile, quantitative medical knowledge was integrated into the training process through a curriculum learning framework, which enables the model to first learn from "easier" samples and then transition to "harder" samples to reach better performance. In addition, our multiview network can work both in a dual-view setting and with a single view as input. We evaluate our method through extensive experiments on a classification task of elbow fracture with a dataset of 1,964 images. Results show that our method outperforms two related methods on bone fracture study in multiple settings, and our technique is able to boost the performance of the compared methods. The code is available at https://github.com/ljaiverson/multiview-curriculum.

Deep learning evaluation of pelvic radiographs for position, hardware presence, and fracture detection.

PURPOSE: Recent papers have shown the utility of deep learning in detecting hip fractures with pelvic radiographs, but there is a paucity of research utilizing deep learning to detect pelvic and acetabular fractures. Creating deep learning models also requires appropriately labeling x-ray positions and hardware presence. Our purpose is to train and test deep learning models to detect pelvic radiograph position, hardware presence, and pelvic and acetabular fractures in addition to hip fractures. MATERIAL AND METHODS: Data was retrospectively acquired between 8/2009-6/2019. A subset of the data was split into 4 position labels and 2 hardware labels to create position labeling and hardware detecting models. The remaining data was parsed with these trained models, labeled based on 6 "separate" fracture patterns, and various fracture detecting models were created. A receiver operator characteristic (ROC) curve, area under the curve (AUC), and other output metrics were evaluated. RESULTS: The position and hardware models performed well with AUC of 0.99-1.00. The AUC for proximal femoral fracture detection was as high as 0.95, which was in line with previously published research. Pelvic and acetabular fracture detection performance was as low as 0.70 for the posterior pelvis category and as high as 0.85 for the acetabular category with the "separate" fracture model. CONCLUSION: We successfully created deep learning models that can detect pelvic imaging position, hardware presence, and pelvic and acetabular fractures with AUC loss of only 0.03 for proximal femoral fracture.

Retraining an open-source pneumothorax detecting machine learning algorithm for improved performance to medical images.

PURPOSE: To validate a machine learning model trained on an open source dataset and subsequently optimize it to chest X-rays with large pneumothoraces from our institution. METHODS: The study was retrospective in nature. The open-source chest X-ray (CXR8) dataset was dichotomized to cases with pneumothorax (PTX) and all other cases (non-PTX), resulting in 41,946 non-PTX and 4696 PTX cases for the training set and 11,120 non-PTX and 541 PTX cases for the validation set. A limited supervision machine learning model was constructed to incorporate both localized and unlocalized pathology. Cases were then queried from our health system from 2013 to 2017. A total of 159 pneumothorax and 682 non-pneumothorax cases were available for the training set. For the validation set, 48 pneumothorax and 1287 non-pneumothorax cases were available. The model was trained, a receiver operator curve (ROC) was created, and output metrics, including area under the curve (AUC), sensitivity and specificity were calculated. RESULTS: Initial training of the model using the CXR8 dataset resulted in an AUC of 0.90 for pneumothorax detection. Naively inferring our own validation dataset on the CXR8 trained model output an AUC of 0.59. After re-training the model with our own training dataset, the validation dataset inference output an AUC of 0.90. CONCLUSION: Our study showed that even though you may get great results on open-source datasets, those models may not translate well to real world data without an intervening retraining process.

Ankle Fracture Detection Utilizing a Convolutional Neural Network Ensemble Implemented with a Small Sample, De Novo Training, and Multiview Incorporation.

To determine whether we could train convolutional neural network (CNN) models de novo with a small dataset, a total of 596 normal and abnormal ankle cases were collected and processed. Single- and multiview models were created to determine the effect of multiple views. Data augmentation was performed during training. The Inception V3, Resnet, and Xception convolutional neural networks were constructed utilizing the Python programming language with Tensorflow as the framework. Training was performed using single radiographic views. Measured output metrics were accuracy, positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity. Model outputs were evaluated using both one and three radiographic views. Ensembles were created from a combination of CNNs after training. A voting method was implemented to consolidate the output from the three views and model ensemble. For single radiographic views, the ensemble of all 5 models produced the best accuracy at 76%. When all three views for a single case were utilized, the ensemble of all models resulted in the best output metrics with an accuracy of 81%. Despite our small dataset size, by utilizing an ensemble of models and 3 views for each case, we achieved an accuracy of 81%, which was in line with the accuracy of other models using a much higher number of cases with pre-trained models and models which implemented manual feature extraction.

3-Dimensional Printed Models May Be a Useful Tool When Planning Revision Anterior Cruciate Ligament Reconstruction.

PURPOSE: To determine whether using 3-dimensional (3D)-printed models in addition to computed tomography (CT) scans to evaluate the primary femoral and tibial tunnels before revision anterior cruciate ligament (ACL) reconstruction leads to better agreement with the surgical approach than CT alone. METHODS: Fifteen patients who underwent revision ACL reconstruction were retrospectively identified. The mean age was 24.3 years, and 73% were female. Using only CT images, 3 board-certified orthopaedists and 5 sports medicine orthopaedic fellows evaluated whether the existing tibial and femoral tunnels were acceptable for the revision surgery. Subsequently, 3D-printed models were made available in addition to the CT scan, and the same questions were asked. RESULTS: For the attending orthopaedic physicians, adding the 3D-printed models did not have a significant impact on the tibial or femoral tunnel agreement compared with the surgical approach. With the fellow physicians, however, using the 3D-printed models with tibial tunnel evaluation led to a higher agreement rate (76%) compared with CT images alone (63%) (P = .050). Furthermore, with the fellow physicians, there was a higher overall agreement when evaluating both the tibial and femoral tunnels with the addition of 3D-printed models (74%) compared with CT alone (65%) (P = .049). CONCLUSION: Our hypothesis that using 3D-printed models leads to better agreement with the surgical approach was unsupported based on the response of the board-certified orthopaedists. Based on the fellow response, it stands to reason that 3D-printed models may be a useful tool in understanding spatial orientation when planning for revision ACL surgery. LEVEL OF EVIDENCE: IV, retrospective case series.

Hypoxic-ischemic injury: utility of susceptibility-weighted imaging.

Magnetic resonance imaging is increasingly used to assess neonatal hypoxic-ischemic injury, and several scoring systems were developed to predict neurologic outcomes in these patients. We examined the magnetic resonance imaging studies of 33 neonates/infants who manifested acute perinatal hypoxic-ischemic injuries. Using a seven-point susceptibility-weighted imaging categorical grading scale, each patient received a "prominence of vein" score, which was dichotomized into a "normal" or "abnormal" group. Six-month outcomes were assessed using the Pediatric Cerebral Performance Category Scale. We then determined whether "prominence of vein" scores correlated with neurologic outcomes in patients with hypoxic-ischemic injuries, and compared these results with the Barkovich magnetic resonance imaging scoring system. Patients with "normal" "prominence of vein" scores demonstrated better outcomes (mean Pediatric Cerebral Performance Category Scale value = 2) than patients with "abnormal" "prominence of vein" scores (mean Pediatric Cerebral Performance Category Scale value = 4). The dichotomized "prominence of vein" groups demonstrated correlations with the Barkovich magnetic resonance imaging scores of the proton density-weighted basal ganglia, watershed, and combined basal ganglia/watershed regions. The susceptibility-weighted imaging categorical grading scale may aid in predicting neurologic outcomes after hypoxic-ischemic injuries.

Chronic actinic prurigo presenting in an adopted child.

In diagnosing actinic prurigo (AP), the patients' ethnic background is very helpful as this condition is associated with very specific ethnic groups. We discuss a patient with an unknown family history who presented with a rash that initially seemed like lupus, but was subsequently diagnosed as AP upon further evaluations.

Improved neurological function in a paediatric patient following Onyx embolization of a cervical glomus arteriovenous malformation.

Spinal cord intramedullary arteriovenous malformations (AVMs) pose a therapeutic challenge. Because of their complex angioanatomy, surgical excision of these lesions is difficult at best. Over the past decade, endovascular therapy has been established as an alternative treatment modality. As an embolic agent, N-butyl cyanoacrylate (NBCA) posed several problems such as difficulty of use and unpredictable performance. Onyx (ev3, Irvine, California, USA), an alternative liquid embolic agent, possesses several advantageous properties, such as increased control of agent delivery, over previous embolic agents like NBCA. However, reports of Onyx use in treating spinal intramedullary AVMs are still rare, especially in paediatric patients. We report a paediatric patient with glomus-type spinal intramedullary AVM treated successfully with Onyx with intermediate-term outcome.

A case of tuberculosis in a patient on Efalizumab and Etanercept for treatment of refractory palmopustular psoriasis and psoriatic arthritis.

We discuss a patient with a history of a positive tuberculin skin test, who presented with severe, recalcitrant palmoplantar pustular psoriasis with psoriatic arthritis whose symptoms did not resolve with monotherapy of etanercept (Enbrel) or efalizumab (Raptiva) alone, but did respond to a combination of both biologics. However, our patient was later found to have re-activation tuberculosis after long-term treatment. This case highlights many key points for treatment of psoriasis and psoriatic arthritis with biologics. Namely, that recalcitrant psoriatic skin lesions may have good clearing on one biologic, such as efalizumab, and arthritic symptoms can be well-controlled with etanercept, leading patients to be on two different biologics concurrently to control symptoms. However, it also highlights the importance of determining a patient's tuberculosis status, initiating prophylactic anti-tuberculosis therapy prior to starting treatment with etanercept, and setting up an adequate treatment regime if the patient develops active tuberculosis during therapy with etanercept.