Deep learning classification of pediatric spinal radiographs for use in large scale imaging registries.

PURPOSE: The purpose of this study is to develop and apply an algorithm that automatically classifies spine radiographs of pediatric scoliosis patients. METHODS: Anterior-posterior (AP) and lateral spine radiographs were extracted from the institutional picture archive for patients with scoliosis. Overall, there were 7777 AP images and 5621 lateral images. Radiographs were manually classified into ten categories: two preoperative and three postoperative categories each for AP and lateral images. The images were split into training, validation, and testing sets (70:15:15 proportional split). A deep learning classifier using the EfficientNet B6 architecture was trained on the spine training set. Hyperparameters and model architecture were tuned against the performance of the models in the validation set. RESULTS: The trained classifiers had an overall accuracy on the test set of 1.00 on 1166 AP images and 1.00 on 843 lateral images. Precision ranged from 0.98 to 1.00 in the AP images, and from 0.91 to 1.00 on the lateral images. Lower performance was observed on classes with fewer than 100 images in the dataset. Final performance metrics were calculated on the assigned test set, including accuracy, precision, recall, and F1 score (the harmonic mean of precision and recall). CONCLUSIONS: A deep learning convolutional neural network classifier was trained to a high degree of accuracy to distinguish between 10 categories pre- and postoperative spine radiographs of patients with scoliosis. Observed performance was higher in more prevalent categories. These models represent an important step in developing an automatic system for data ingestion into large, labeled imaging registries.

Automated measurements of interscrew angles in vertebral body tethering patients with deep learning.

BACKGROUND CONTEXT: Vertebral body tethering is the most popular nonfusion treatment for adolescent idiopathic scoliosis. The effect of the tether cord on the spine can be segmentally assessed by comparing the angle between two adjacent screws (interscrew angle) over time. Tether breakage has historically been assessed radiographically by a change in adjacent interscrew angle by greater than 5° between two sets of imaging. A threshold for growth modulation has not yet been established in the literature. These angle measurements are time consuming and prone to interobserver variability. PURPOSE: The purpose of this study was to develop an automated deep learning algorithm for measuring the interscrew angle following VBT surgery. STUDY DESIGN/SETTING: Single institution analysis of medical images. PATIENT SAMPLE: We analyzed 229 standing or bending AP or PA radiographs from 100 patients who had undergone VBT at our institution. OUTCOME MEASURES: Physiologic Measures: An image processing algorithm was used to measure interscrew angles. METHODS: A total of 229 standing or bending AP or PA radiographs from 100 VBT patients with vertebral body tethers were identified. Vertebral body screws were segmented by hand for all images and interscrew angles measured manually for 60 of the included images. A U-Net deep learning model was developed to automatically segment the vertebral body screws. Screw label maps were used to develop and tune an image processing algorithm which measures interscrew angles. Finally, the completed model and algorithm pipeline was tested on a 30-image test set. Dice score and absolute error were used to measure performance. RESULTS: Inter- and Intra-rater reliability for manual angle measurements were assessed with ICC and were both 0.99. The segmentation model Dice score against manually segmented ground truth across the 30-image test set was 0.96. The average interscrew angle absolute error between the algorithm and manually measured ground truth was 0.66° and ranged from 0° to 2.67° in non-overlapping screws (N=206). The primary modes of failure for the model were overlapping screws on a right thoracic/left lumbar construct with two screws in one vertebra and overexposed images. An algorithm step which determines whether an overlapping screw was present correctly identified all overlapping screws, with no false positives. CONCLUSION: We developed and validated an algorithm which measures interscrew angles for radiographs of vertebral body tether patients with an accuracy of within 1° for the majority of interscrew angles. The algorithm can process five images per second on a standard computer, leading to substantial time savings. This algorithm may be used for rapid processing of large radiographic databases of tether patients and could enable more rigorous definitions of growth modulation and cord breakage to be established.

Comparison of peripheral leukocyte parameters in patients receiving conventionally and hypofractionated radiotherapy schemes for the treatment of newly diagnosed glioblastoma.

Introduction: Treatment for glioblastomas, aggressive and nearly uniformly fatal brain tumors, provide limited long-term success. Immunosuppression by myeloid cells in both the tumor microenvironment and systemic circulation are believed to contribute to this treatment resistance. Standard multi-modality therapy includes conventionally fractionated radiotherapy over 6 weeks; however, hypofractionated radiotherapy over 3 weeks or less may be appropriate for older patients or populations with poor performance status. Lymphocyte concentration changes have been reported in patients with glioblastoma; however, monocytes are likely a key cell type contributing to immunosuppression in glioblastoma. Peripheral monocyte concentration changes in patients receiving commonly employed radiation fractionation schemes are unknown. Methods: To determine the effect of conventionally fractionated and hypofractionated radiotherapy on complete blood cell leukocyte parameters, retrospective longitudinal concentrations were compared prior to, during, and following standard chemoradiation treatment. Results: This study is the first to report increased monocyte concentrations and decreased lymphocyte concentrations in patients treated with conventionally fractionated radiotherapy compared to hypofractionated radiotherapy. Discussion: Understanding the impact of fractionation on peripheral blood leukocytes is important to inform selection of dose fractionation schemes for patients receiving radiotherapy.

Few-shot biomedical image segmentation using diffusion models: Beyond image generation.

BACKGROUND: Medical image analysis pipelines often involve segmentation, which requires a large amount of annotated training data, which is time-consuming and costly. To address this issue, we proposed leveraging generative models to achieve few-shot image segmentation. METHODS: We trained a denoising diffusion probabilistic model (DDPM) on 480,407 pelvis radiographs to generate 256 âœ• 256 px synthetic images. The DDPM was conditioned on demographic and radiologic characteristics and was rigorously validated by domain experts and objective image quality metrics (Frechet inception distance [FID] and inception score [IS]). For the next step, three landmarks (greater trochanter [GT], lesser trochanter [LT], and obturator foramen [OF]) were annotated on 45 real-patient radiographs; 25 for training and 20 for testing. To extract features, each image was passed through the pre-trained DDPM at three timesteps and for each pass, features from specific blocks were extracted. The features were concatenated with the real image to form an image with 4225 channels. The feature-set was broken into random patches, which were fed to a U-Net. Dice Similarity Coefficient (DSC) was used to compare the performance with a vanilla U-Net trained on radiographs. RESULTS: Expert accuracy was 57.5 % in determining real versus generated images, while the model reached an FID = 7.2 and IS = 210. The segmentation UNet trained on the 20 feature-sets achieved a DSC of 0.90, 0.84, and 0.61 for OF, GT, and LT segmentation, respectively, which was at least 0.30 points higher than the naively trained model. CONCLUSION: We demonstrated the applicability of DDPMs as feature extractors, facilitating medical image segmentation with few annotated samples.

Probabilistic tractography of the extracranial branches of the trigeminal nerve using diffusion tensor imaging.

PURPOSE: The peripheral course of the trigeminal nerves is complex and spans multiple bony foramen and tissue compartments throughout the face. Diffusion tensor imaging of these nerves is difficult due to the complex tissue interfaces and relatively low MR signal. The purpose of this work is to develop a method for reliable diffusion tensor imaging-based fiber tracking of the peripheral branches of the trigeminal nerve. METHODS: We prospectively acquired imaging data from six healthy adult participants with a 3.0-Tesla system, including T2-weighted short tau inversion recovery with variable flip angle (T2-STIR-SPACE) and readout segmented echo planar diffusion weighted imaging sequences. Probabilistic tractography of the ophthalmic, infraorbital, lingual, and inferior alveolar nerves was performed manually and assessed by two observers who determined whether the fiber tracts reached defined anatomical landmarks using the T2-STIR-SPACE volume. RESULTS: All nerves in all subjects were tracked beyond the trigeminal ganglion. Tracts in the inferior alveolar and ophthalmic nerve exhibited the strongest signal and most consistently reached the most distal landmark (58% and 67%, respectively). All tracts of the inferior alveolar and ophthalmic nerve extended beyond their respective third benchmarks. Tracts of the infraorbital nerve and lingual nerve were comparably lower-signal and did not consistently reach the furthest benchmarks (9% and 17%, respectively). CONCLUSION: This work demonstrates a method for consistently identifying and tracking the major nerve branches of the trigeminal nerve with diffusion tensor imaging.

A Deep Learning Tool for Automated Landmark Annotation on Hip and Pelvis Radiographs.

BACKGROUND: Automatic methods for labeling and segmenting pelvis structures can improve the efficiency of clinical and research workflows and reduce the variability introduced with manual labeling. The purpose of this study was to develop a single deep learning model to annotate certain anatomical structures and landmarks on antero-posterior (AP) pelvis radiographs. METHODS: A total of 1,100 AP pelvis radiographs were manually annotated by 3 reviewers. These images included a mix of preoperative and postoperative images as well as a mix of AP pelvis and hip images. A convolutional neural network was trained to segment 22 different structures (7 points, 6 lines, and 9 shapes). Dice score, which measures overlap between model output and ground truth, was calculated for the shapes and lines structures. Euclidean distance error was calculated for point structures. RESULTS: Dice score averaged across all images in the test set was 0.88 and 0.80 for the shape and line structures, respectively. For the 7-point structures, average distance between real and automated annotations ranged from 1.9 mm to 5.6 mm, with all averages falling below 3.1 mm except for the structure labeling the center of the sacrococcygeal junction, where performance was low for both human and machine-produced labels. Blinded qualitative evaluation of human and machine produced segmentations did not reveal any drastic decrease in performance of the automatic method. CONCLUSION: We present a deep learning model for automated annotation of pelvis radiographs that flexibly handles a variety of views, contrasts, and operative statuses for 22 structures and landmarks.

Effect of lesion temperature on the durability of percutaneous radiofrequency rhizotomies to treat trigeminal neuralgia.

OBJECTIVE: Percutaneous radiofrequency rhizotomy is a common procedure for trigeminal neuralgia (TN) that creates thermocoagulative lesions in the trigeminal ganglion. Lesioning parameters for the procedure are left to the individual surgeon's discretion, and published guidance is primarily anecdotal. The purpose of this work was to assess the role of lesioning temperature on long-term surgical outcomes. METHODS: This was a retrospective analysis of patients who underwent percutaneous radiofrequency rhizotomy from 2009 to 2020. Patient data, including demographics, disease presentation, surgical treatment, and outcomes, were collected from medical records. The primary endpoint was the recurrence of TN pain. Univariate and multivariate Cox proportional hazards regressions were used to assess the impact of chosen covariates on pain-free survival. RESULTS: A total of 280 patients who had undergone 464 procedures were included in the analysis. Overall, roughly 80% of patients who underwent rhizotomy would have a recurrence within 10 years. Lower lesion temperature was predictive of longer periods without pain recurrence (HR 1.05, p < 0.001). The inclusion of lesion time, postoperative numbness, prior history of radiofrequency rhizotomy, surgeon, and multiple sclerosis as confounding variables did not affect the hazard ratio or the statistical significance of this finding. Postoperative numbness and the absence of multiple sclerosis were significant protective factors in the model. CONCLUSIONS: The study findings suggest that lower lesion temperatures and, separately, postoperative numbness result in improved long-term outcomes for patients with TN who undergo percutaneous radiofrequency rhizotomies. Given the limitations of retrospective analysis, the authors suggest that a prospective multisite clinical trial testing lesion temperatures would provide definitive guidance on this issue with specific recommendations about the number needed to treat and trial design.

The practical limits of high-quality magnetic resonance imaging for the diagnosis and classification of trigeminal neuralgia.

BACKGROUND: Neurovascular compression (NVC) has been the primary hypothesis for the underlying mechanism of classical trigeminal neuralgia (TN). However, a substantial body of literature has emerged highlighting notable exceptions to this hypothesis. The purpose of this study is to assess the reliability and diagnostic accuracy of high resolution, high contrast MRI-determined neurovascular contact for TN. METHODS: We performed a retrospective, randomized, and blinded parallel characterization of neurovascular interaction and diagnosis in a population of TN patients and controls using four expert reviewers. Performance statistics were calculated, as well as assessments for generalizability using shuffled bootstraps. RESULTS: Fair to moderate agreement (ICC: 0.32-0.68) about diagnosis between reviewers was observed using MRIs from 47 TN patients and 47 controls. On average reviewers performed no better than chance when diagnosing participants, with an accuracy of 0.57 (95% CI 0.40, 0.59) per patient. CONCLUSION: While MRI is useful in determining structural causes in secondary TN, expert reviewers do no better to only slightly better than chance with distinguishing TN with MRI, despite moderate agreement. Further, the causal role of NVC for TN is not clear, limiting the applicability of MRI to diagnose or prognosticate treatment of TN.

Predicting Glioblastoma Cellular Motility from In Vivo MRI with a Radiomics Based Regression Model.

Characterizing the motile properties of glioblastoma tumor cells could provide a useful way to predict the spread of tumors and to tailor the therapeutic approach. Radiomics has emerged as a diagnostic tool in the classification of tumor grade, stage, and prognosis. The purpose of this work is to examine the potential of radiomics to predict the motility of glioblastoma cells. Tissue specimens were obtained from 31 patients undergoing surgical resection of glioblastoma. Mean tumor cell motility was calculated from time-lapse videos of specimen cells. Manual segmentation was used to define the border of the enhancing tumor T1-weighted MR images, and 107 radiomics features were extracted from the normalized image volumes. Model parameter coefficients were estimated using the adaptive lasso technique validated with leave-one-out cross validation (LOOCV) and permutation tests. The R-squared value for the predictive model was 0.60 with p-values for each individual parameter estimate less than 0.0001. Permutation test models trained with scrambled motility failed to produce a model that out-performed the model trained on the true data. The results of this work suggest that it is possible for a quantitative MRI feature-based regression model to non-invasively predict the cellular motility of glioblastomas.

Identifying symptomatic trigeminal nerves from MRI in a cohort of trigeminal neuralgia patients using radiomics.

INTRODUCTION: Trigeminal neuralgia (TN) is a devastating neuropathic condition. This work tests whether radiomics features derived from MRI of the trigeminal nerve can distinguish between TN-afflicted and pain-free nerves. METHODS: 3D T1- and T2-weighted 1.5-Tesla MRI volumes were retrospectively acquired for patients undergoing stereotactic radiosurgery to treat TN. A convolutional U-net deep learning network was used to segment the trigeminal nerves from the pons to the ganglion. A total of 216 radiomics features consisting of image texture, shape, and intensity were extracted from each nerve. Within a cross-validation scheme, a random forest feature selection method was used, and a shallow neural network was trained using the selected variables to differentiate between TN-affected and non-affected nerves. Average performance over the validation sets was measured to estimate generalizability. RESULTS: A total of 134 patients (i.e., 268 nerves) were included. The top 16 performing features extracted from the masks were selected for the predictive model. The average validation accuracy was 78%. The validation AUC of the model was 0.83, and sensitivity and specificity were 0.82 and 0.76, respectively. CONCLUSION: Overall, this work suggests that radiomics features from MR imaging of the trigeminal nerves correlate with the presence of pain from TN.

A radiomics-based model for predicting local control of resected brain metastases receiving adjuvant SRS.

PURPOSE: Adjuvant radiosurgery to the cavities of surgically resected brain metastases provides excellent local tumor control while reducing the risk of deleterious cognitive decline associated with whole brain radiotherapy. A subset of these patients, however, will develop disease recurrence following radiosurgery. In this study, we sought to assess the predictive capability of radiomic-based models, as compared with standard clinical features, in predicting local tumor control. METHODS: We performed a retrospective chart review of patients treated with adjuvant radiosurgery for resected brain metastases at the "Institution" from 2009 to 2019. Shape, intensity and texture based radiomics features of the cavities were extracted from the pre-radiosurgery treatment planning MRI scans and trained using a gradient boosting technique with K-fold cross validation. RESULTS: In total, 71 cavities from 67 treated patients were included for analysis. The 6 and 12 month local control estimates were 86% and 76%, respectively. The 6 and 12 month overall survival was 78% and 55%, respectively. Thirty-six patients developed intracranial failures outside of the surgical cavity. The predictive model for local control trained on imaging features from the whole cavity achieved an area-under-the-curve (AUC) of 0.73 on the validation set versus an AUC of 0.40 for the clinical features. CONCLUSIONS: Here we report a single institutional experience using radiomic-based predictive modeling of local tumor control following adjuvant Gamma Knife radiosurgery for resected brain metastases. We found the radiomics features to provide more robust predictive models of local control rates versus clinical features alone. Such techniques could potentially prove useful in the clinical setting and warrant further investigation.