Physical Exam, Static & Dynamic Ultrasound Assessment, & Treatment of Thoracolumbar Fascia (TLF) Mediated Low Back Pain

Prior to developing a randomized trial using sonographic TLF imaging guided hydrodissection or the use of OMT for TLF mediated LBP, the investigators must demonstrate the association of imaging characteristics with TFL mediated LBP. The investigators also wish to data mine US images of the TFL to further develop non-clinical US image assessment methods through textural and machine learning analysis. The investigators will perform two different pilot, prospective studies (single factor, repeated measure design), that will share a common asymptomatic population and similar image assessment (clinical \& research based) and functional assessment evaluation protocols to look for biomarkers of TLF in acute \& chronic LBP. The CLBP patients in each study (different populations) will undergo a pilot treatment arm in AIM 3: 1. Hydrodissection (DoD Study) for the treatment of TLF glide impairment in CLBP 2. OMT (AOA Study) for the treatment of TLF glide impairment in CLBP Both studies will be harnessing the synergistic potential of quantitative imaging (including grayscale ultrasound, shear-wave elastography, and computational approaches such as texture analysis) coupled with artificial intelligence to establish TLF imaging biomarkers critical to assessing the efficacy of OMT or hydrodissection in CLBP. These tools can serve as clinical decision support in the identification, diagnosis, treatment, and monitoring of low back pain. The investigators will characterize the sonographic presentation of the TLF, statically \& dynamically in both study populations: 1. symptomatic (CLBP) and asymptomatic individuals (Aim 1 AOA), and 2. in symptomatic (CLBP \& ALBP) and asymptomatic individuals (Aim 1 DoD) All subjects will be a convenience sample of men and women, ages 18-50 years old. After informed consent is obtained, in both studies, subjects will fill out the following standardized pain-related questionnaires and IRB approved study specific demographic/history questionnaires prior to undergoing US imaging: 1. Study-specific, IRB approved demographic questionnaire for demographics \& medical history; 2. Baeke Physical Activity Questionnaire; 3. Oswestry Low Back Pain Disability Questionnaire; 4. Visual analog pain scale: Wong-Baker FACES Pain Rating Scale Biopsychosocial Assessment: 5. 7-item Generalized Anxiety Disorder assessment (GAD-7) to assess anxiety; 6. 9-item Patient Health Questionnaire (PHQ-9) to assess depression. 7. 8-item Social Functioning Questionnaire (SFQ) to assess social functioning of subjects. All three tools are self-reporting questionnaires that are free, reliable, and established screening tools. These survey instruments will be administered at Baseline (Aim 1), and at 2, 4, 6, 12, and 24 weeks post-treatment (Aim 3: clinical trial). All these instruments are commonly utilized in LBP studies. The demographic questionnaire, completed by the subject, but reviewed by the researcher with the subject, will obtain from all subjects the following information: age, sex, body mass index (BMI), MSK injury or surgery of the spine and/or torso, abdominal surgical history, history of low back injury, history of pregnancy, history of LBP and level of daily physical activity (See Attached Demographic Questionnaire). The demographic questionnaire will only be administered at the beginning of the study after obtaining informed consent. All subjects in both studies will complete the following functional exam assessments 1. Sit and Reach Test (to quantify flexibility of the hamstrings and lower back); Subjects will then undergo Static US imaging of their TLF. The US imaging will be conducted by co-Is Drs. Kozar assisted by a sports medicine fellow (PGY4 or 5) or Osteopathic Neuromusculoskeletal Medicine Resident (PGY4 or 5). Subjects will then undergo dynamic US imaging of their TLF. The US imaging will be conducted by Dr. Kozar assisted by a sports medicine fellow (PGY4 or 5) or Osteopathic Neuromusculoskeletal Medicine Resident (PGY4 or 5). TLF glide will be assessed by ultrasound in three positions: 1. Prone Flexion on a motorized table 2. Prone Active Straight Leg Raise All ultrasound assessments will be administered at Baseline (Aim 1) and at 2, 4, 6, 12, and 24 weeks post-treatment (Aim 3: clinical trial). Ultrasound Imaging Protocol Measurements: Static Image Assessment: * Images were obtained using SuperSonic MACH 30 (Hologic Inc, France) using UltraFast Imaging SonicSoftware with Linear 50mm wide, 5-18 MHz probe. * During each imaging session, grey-scale (GS), power Doppler (PD) \& sheer-wave elastography (SWE) ultrasound modes will be utilized and recorded simultaneously using Supersonic's TriVu software technology to obtain longitudinal and transverse views of 31 probe positions respresenting key transition zones of the subjects TLF (Figure 2). These locations were chosen to best sample multiple key areas of load transfer and functional stabilization within transition zones of different attachments of the TLF. Each probe position will have 1 grey scale (b-mode), 1 power doppler and 2 Shear-Wave Elastography images (after a 5 second equilibration time). In total, The investigators will obtain 124 static images per scanning session per patient. (See figure 2) Clinical US grading criteria will be used to assess all static images: 1. TLF echogenicity (GS images only) 2. TLF Thickness (GS images only) 3. Marginal Blurring (GS images only) 4. Fascial Herniation (GS images only) 5. Fascia Linearity (GS images only) 6. Doppler positivity (PD images only) 7. Stiffness (SWE images only) Textural Analysis \& Machine Learning Analysis will be used to assess all images: 1. 1st order, 2nd order, and higher order (e.g. blob analysis and run length analysis) texture parameters will be calculated for each image 2. Machine learning algorithms will be trained to identify the TLF in each image and distinguish it from adjacent tissues/structures 3. Machine learning algorithms will be trained to extract quantitative features from each image and identify which features and texture parameters can be used to distinguish images from patients with acute vs. chronic pain vs. healthy tissue 4. Machine learning algorithms will be trained to 1) classify the images as chronic vs. acute pain vs. healthy tissue, and 2) predict a subject's post-treatment pain and clinical function scores based on pre-treatment imaging, pain and clinical function scores Ultrasound Imaging Protocol Measurements: Dynamic Image Assessment: ... All dynamic cine loops obtained on GE LOGIQ S8 US, linear Matrix ML6-15 probe @15MHz, with 3 focal points within TLF. Ultrasound cine recordings of posterior TLF layer were obtained bilaterally as described below: 1. Five cycles of 20 degree prone trunk flexion (PTF) induced by a motorized table were recorded, with probe centered vertically over L3-L4 paraspinal muscles. 2. Five cycles of 12" prone active straight leg raise (pASLR), contralateral \& ipsilateral, were recorded with probe centered vertically over T12-L1 paraspinal muscles. A single researcher analyzed cine recordings using Tracker software. Maximum relative movement was calculated in cranio-caudal and anterior-posterior directions, averaged, and reported as mean glide and mean layer separation. A linear mixed effect model was used to compare pain status, sex, age, body side, BMI, and movement type. T-tests were used to evaluate mean/maximum glide comparisons. All TLF dynamic ultrasound assessments will be administered at Baseline (Aim 1) and at 2, 4, 6, 12, and 24 weeks post-treatment (Aim 3: clinical trial). Those subjects that enter AIM 3 will also have the following clinical exam assessments: Common clinical LB exam maneuvers consisting of: * Supine: Straight Leg Raise (SLR - passive), Thigh Thrust (TT), Active Straight Leg Raise (ASLRs), and Duel Legged Lowering (DLL) * Prone: Hip Extension (HE), Youmans Test (Y), Passive Lumbar Extension (PLE - via each leg), Prone Press Up (PPU - active lumbar extension), and Active Straight Leg Raise (ASLRp) w \& w/o TLF Pre-Stress; Prone Hip Extension Motor Firing Pattern of the Deep Sacral Gluteus Maximus (DSG), Prone Hip Extension Strength (pHES) w \& s/out LIFT/LIST Pre-Stress, Prone Hip Eternal Rotation Strength (pHERS) w \& s/out LIFT/LIST Pre-Stress. * AOA Study only: Osteopathic structural assessment, consisting of area of greatest restriction screening exam, followed by segmental diagnosis of somatic dysfunction in 3-5 regions. All subjects will have clinical exam assessments at Baseline. Those subjects entering AIM 3 will have clinical exam assessments at each treatment session (x3) and at 2, 4, 6, 12, and 24 weeks post-treatment. As this is a pilot study, the investigators will not be blinded to measures. 2.6 Data Analysis: Statistical Analyses The investigators will summarize the quantitative characteristics of the images using principal components to reduce dimensionality and create independent orthogonal summaries from the images. Using the principal component, the investigators will conduct Analysis of Variance (ANOVA) testing to determine whether the mean value (e.g., each of the texture parameters) differs by pain group. Furthermore, the investigators will estimate correlations of quantitative US texture parameters across modalities (GS, SWE, and power Doppler). The investigators will summarize these associations using Pearson's correlation coefficient and, due to the repeated measurements taken within an individual, 95% confidence intervals using bootstrapping. The investigators will also test for associations, using principal components regressions, between clinical assessments and quantitative US texture parameters. Data interpretation The investigators expect that our detailed image texture analysis of three types of US images will reveal novel biomarkers which distinguish individuals who are present with acute, chronic or no pain. Statistical analyses of our comprehensive dataset (Table 1) will advance knowledge of how regional imaging features are potentially associated with pain and function. Specific Aim 2: Develop a machine learning (ML) classification algorithm to reliably distinguish abnormal myofascial tissue in acute versus chronic pain stages from healthy tissue. Utilizing the datasets acquired in Specific Aim 1, the investigators will automate the identification of the TLF from surrounding tissues in US images using image segmentation. The investigators will then develop feature selection techniques associated with acute and chronic stages of myofascial pain. Lastly, The investigators will train a ML algorithm to classify (Random Forest and Speeded-Up Robust Features (SURF) algorithms) and identify US image characteristics that are associated with acute and chronic stages of myofascial pain. Lastly, the investigators will train a ML algorithm to classify US images of the TLF into three classes: 1) acute pain, 2) chronic pain, and 3) healthy tissue. The investigators seek to attain a minimum classification accuracy of 90%. All of the proposed analyses in Aim 2 will be based upon de-identified ultrasound images. Image Segmentation To automate the identification and isolation of the TLF in US images, The investigators will use both commercial segmentation software as well as deep learning segmentation algorithms. With the MATLAB Image Segmenter App, the investigators will apply k-means clustering and region growing segmentation to identify the border of the TLF and to identify hyperechoic and hypoechoic regions/blobs within each image. The investigators will also use transfer learning to adapt fully convolutional networks (FCNs) that have been trained for segmenting natural images. The investigators will train these FCNs using the PyTorch deep learning library for Python and a subset of TLF images manually segmented at the pixel-level by a trained clinician. Results of each algorithm will be validated by manual identification of the TLF by the trained clinician. For each image, the TLF border identified by the segmentation algorithm will be used in the QUS textural analysis technique in lieu of manually defining the ROI for pixel-wise analysis. The investigators will evaluate the quality of the segmentation by measuring the intersection-over-union (IoU) metric, which compares the portion of the image identified as the ROI by both the clinician and the algorithm. The investigators goal is to identify an algorithm that reaches at least 75% IoU, which is just below the highest IoU scores achieved with large natural image datasets. Feature Selection The Random Forest (RF) algorithm will be applied to the image feature library obtained from the images acquired in Aim 1 to determine the set of image characteristics which are biomarkers for 1) acute myofascial pain and 2) chronic myofascial pain, compared with healthy tissue. The feature library will consist of the 7 clinical grading criteria and 12 image texture features obtained at 9 anatomical locations and 2 probe orientations, as well as image features identified via the Speeded-Up Robust Features (SURF) algorithm. Our recent study successfully used SURF to identify "interest points" (e.g. corners, edges, and blobs) within tendon images, which were then used in a support vector machine (SVM) classification algorithm to distinguish images of tendinopathic tendons from images of healthy tendons with a 77.5% accuracy. Images will be divided into the training and test sets, with two-thirds of the images allocated to training and one-third allocated to test. Features (clinical grading, texture parameters, and SURF features) will be extracted from the images. The RF algorithm will then train multiple decision trees using the training set. Each tree will be trained on a randomly selected subset of the features. The RF algorithm will train trees to classify the images as one of three classes: 1) healthy tissue, 2) chronic myofascial pain, or 3) acute myofascial pain. The RF algorithm then provides a ranking of how frequently features are used in the trained trees. The feature ranking will be used to determine the top 10 features that are indicative of each class, and thus can serve as biomarkers. The trained RF algorithm will then be run on the test set of images, and the prediction accuracy of the model will be used to assess the effectiveness of the selected features. To measure the effectiveness of feature selection, the investigators will compare the test accuracy of classifiers using all features to classifiers using only the selected features. If the selected features are indeed the most relevant, the accuracy of a model trained on them should be similar or better than that of a model trained on the full feature set. Our target is to achieve at least 90% of the accuracy of a model trained without feature selection. Image Classification The investigators will first update our two existing ML classification algorithms (SVM and CNN) to reliably distinguish abnormal myofascial tissue in chronic vs. acute pain stages from healthy tissue using the GS, SWE, and power Doppler images of Aim 1. Images will be divided into training and test sets, with two-thirds of the images allocated to training and one-third allocated to test, and all images from each patient maintained within the same set. The training set will be labeled as healthy tissue, chronic myofascial pain, or acute myofascial pain based on clinical assessment. The CNN will train directly on the training set of images, while for the SVM, the images will be pre-processed using SURF to extract image features (e.g. corners, edges and blobs). The SVM will then train on these extracted features. The accuracy, specificity, and sensitivity of each algorithm will be reported as a confusion matrix by comparing the algorithm's classification of the test images with their true class. The investigators will then improve the accuracy of the SVM algorithm through novel ML techniques and incorporation of the features identified through our feature selection algorithm. To optimize our classification algorithms, the investigators will continue investigating additional classical ML and deep learning methods. Recent results have demonstrated the power of neural attention models for semantic segmentation. Additionally, the investigators will use weakly supervised learning methods to incorporate the features identified through our feature selection algorithms into the learning process. These weakly supervised learning methods will help address the ML challenge of training with limited labeled examples, a problem common in many medical imaging tasks. Specific Aim 3: Assess the therapeutic efficacy of hydrodissection as a novel treatment for TLF pain using quantitative US imaging and ML tools. The investigators will perform a pilot, prospective study of percutaneous, ultrasound guided hydrodissection as a treatment effect over time. Subjects with chronic LBP will undergo US imaging and functional assessments as described in Aim 1 before and after hydrodissection of the TLF at 2 locations bilaterally. The investigators will use the trained SVM and CNN algorithms developed in Aim 2 to classify longitudinal US images acquired following hydrodissection treatment into healthy, acute, or chronic pain categories. The investigators hypothesize that patients treated with hydrodissection will exhibit a reduction in LBP, restoration of functional capability, and US imaging biomarkers consistent with a pain-free state (versus acute or chronic pain). Clinically, co-Is Kozar and Brolinson have utilized hydrodissection techniques for the treatment of tendon, nerve, and back pain for several years. Specifically, hydrodissection of fascial tissues such as the iliotibial (IT) band and TLF appear to provide significant and long-term relief of pain syndromes (unpublished clinical experience of the co-Is). The literature on hydrodissection of tendons (tendon scraping) and nerve hydrodissection has been rapidly expanding in recent years; however, little data exists on the use of hydrodissection as treatment for low back pain. Hydrodissection offers an advantage for patients who may not be responsive to manual therapies (massage therapy and spinal manipulation) as well as standard physical therapy approaches. The proposed study would be among the first to examine the effect of hydrodissection on TLF and LBP. ML and Statistical Analyses The investigators will use the trained SVM and CNN algorithms developed in Aim 2 to classify de-identified, longitudinal US images acquired following hydrodissection treatment into CLBP, ALBP, and WLBP categories. The investigators will compare this prediction to the subjects' true class based on clinical assessment and determine the accuracy, specificity, and sensitivity of our algorithm in monitoring response to hydrodissection treatment. The investigators will also train separate SVM and CNN models to predict a subject's post-treatment pain and clinical function scores based on pre-treatment imaging, pain and clinical function scores. This strategy will enable development of a ML tool that can predict a patient's response to hydrodissection prior to treatment. The investigators will utilize linear mixed models to test for changes over time in pain scores and imaging from baseline to 24 weeks. The investigators will program all analyses using SAS 9.4 and R analysis software (4.1.2), saving and documenting all syntax to allow reproduction and replication of all results. It is our expectation that the data from this pilot investigation will demonstrate the association of imaging characteristics with LBP, provide preliminary evidence of pain relief from hydrodissection, and estimate key variance components that are all needed to develop a well powered and justified randomized trial. The investigators will summarize all demographics and clinical outcome variables using descriptive measures of central tendency and dispersion. For continuous variables, the investigators will measure central location using sample mean and dispersion by standard deviation. For categorical variables, the investigators will summarize using proportions. For continuous outcomes such pain scores, reach test, as well as quantified imaging data, linear mixed models will be used to measure treatment effect over time. Careful attention will be paid to the normality assumption which will be examined using histograms and normal probability plots. If the normality assumption appears violated or outliers are present, the investigators will apply transformations. Cohen's D will be calculated for all outcomes to quantify treatment effect size and to assist with design of future studies. The investigators will use the psychosocial survey tools as a gauge: * 1st test would help establish baseline (pre-test) * After testing surveys aim to gauge if subjects show improvement in anxiety, depression, and social functioning as they (hopefully) see improvement in their biological treatments * The demographic questionnaire asks one question of each subject, simply asking if they already see/utilize a mental health therapist (psychiatrist, psychologist, or counselor). The investigators will assume these subjects would have been or are being taught tools to better manage any anxiety, depression, or other social functioning. The investigators will make a comparison of survey responses between those with pre-therapy to those without pre-therapy if the numbers are adequate.