RPS 110 - AI, radiomics and other technologies supporting MSK diagnostics

Author Block: R. Hahnfeldt¹, R. A. Terzis¹, T. M. Dratsch¹, J. Bremm¹, P. Rauen¹, K. Weiss², D. Maintz¹, G. Bratke¹, A-I. Iuga¹; ¹Cologne/DE, ²Hamburg/DE
Purpose: The aim of this study was to investigate whether a 2D MRI lumbar spine protocol with an AI-based SuperResolution reconstruction method meets the requirements for clinical diagnostic purposes.
Methods or Background: In this retrospective study, 25 patients underwent MRI examinations of the lumbar spine using a 1,5T MRI scanner (Philips Ingenia 1.5T, Best, NL). The MRI protocol included three sagittal sequences (STIR, T1 TSE, T2 TSE), and an axial T2 TSE sequence. The images were acquired in both standard and low resolution. Both the clinical standard (Compressed SENSE (CS)) and the new AI-based SuperResolution reconstruction method (SuperRes-AI) were applied. Four experienced readers (two radiologists and two orthopedic surgeons) evaluated the sequences for pathologies (bone marrow edema, neuroforaminal stenosis, disc herniation).
Results or Findings: The acquisition time for the clinical standard sequences was 11 minutes and 5 seconds. In contrast, the acquisition time for the low resolution SuperRes-AI sequences was 7 minutes and 37 seconds (31% scan time reduction). A generalized estimating equations (GEE) analysis revealed no significant differences in the sensitivity for detecting edema between reader groups and reconstruction algorithms (all p>0.99). Bonferroni-corrected post-hoc tests in a GEE analysis revealed significantly higher sensitivity for detecting neuroforaminal stenosis with AI-powered reconstruction compared to conventional algorithms among radiologists (p=0.001), with no other significant differences observed.
Conclusion: The new AI-based SuperResolution reconstruction of low-resolution 2D MRI sequences of the lumbar spine allows for a reduction in acquisition time of approximately 31% without compromising diagnostic quality, showing significantly higher sensitivity for detecting neuroforaminal stenosis. The AI-based SuperResolution method improves MRI efficiency by significantly reducing scan times without compromising image quality, potentially enhancing sensitivity in pathology detection, offering advantages for patient comfort and clinical workflow.
Limitations: Not applicable.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The ethics committee notification can be found under the number DRKS00024156

Author Block: M. Kim¹, J. Song¹, K. Sung², E. Oh¹; ¹Seoul/KR, ²Los Angeles, CA/US
Purpose: The purpose of this study is to develop a deep learning-based system to detect postoperative abnormalities in spine radiographs following lumbar fusion surgery. This system aims to assist radiologists by detecting postoperative abnormalities.
Methods or Background: A total of 1,505 spine radiographs from 85 patients who underwent lumbar fusion surgery were collected at a secondary healthcare facility between February 2018 and January 2022. These radiographs, taken post-operation and during follow-up visits, included anteroposterior, lateral, flexion, and extension views. Annotations for periprosthetic loosening, cage subsidence, and compression fracture were performed by a musculoskeletal radiologist, and verified with CT scans. The Co-DETR model was trained on a subset of 634 radiographs from 74 patients with 726 annotations. The class distribution included 58, 24, and 17 patients yielding 278, 215, and 168 images respectively, with each image averaging 1.10 annotations. Initial training was conducted on a public dataset (FracAtlas), followed by transfer learning to enhance detection of postoperative abnormalities. Negative samples were included to boost training efficiency, and model performance was evaluated using mean Average Precision (mAP).
Results or Findings: Periprosthetic loosening achieved an mAP score of 0.601 with 0.5 IoU threshold. The mAP score for each class of periprosthetic loosening, cage subsidence, and compression fracture were 0.565, 0.667, 0.572, respectively.
Conclusion: The study demonstrates the potential of detecting postoperative abnormalities in spine radiographs after lumbar fusion surgery using deep learning. The results indicate a foundational potential for enhancing diagnostic capabilities in clinical settings. The potential of this approach to improve early detection of complications could lead to more timely interventions and better patient outcomes.
Limitations: Further validation is required to optimize its performance, particularly to support radiologists in settings with limited access to specialists.
Funding for this study: Not applicable
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB No. 2022-08-018

Author Block: I. C. Pizza¹, M. Pedullà², S. Fusco², F. Serpi², D. Albano³, C. Messina², S. Gitto², L. M. Sconfienza²; ¹Eboli/IT, ²Milan/IT, ³Cefalu'/IT
Purpose: The aim of this study is to predict two-year clinical outcome in patients with lumbar spine arthrodesis using machine learning and radiomics based on post-operative X-rays.
Methods or Background: This retrospective study was performed at a tertiary orthopaedic centre and included 162 patients with lumbar spine arthrodesis, post-operative X-rays available for analysis and minimum follow-up of two years. Clinical follow-up was evaluated at two years using Oswestry Disability Index (ODI): ODI≤20 indicated good clinical outcome (n=90), ODI>20 indicated poor clinical outcome (n=72). All X-rays were manually segmented by drawing rectangular regions of interest including the arthrodesis and one adjacent non-operated vertebra on both proximal and distal sides. Radiomic features were extracted. After feature selection and class balancing, machine learning (three ensembles of Random Forest classifiers) was trained, validated using nested 10-fold cross-validation and tested.
Results or Findings: After training and cross-validation, in the test dataset machine learning showed ROC-AUC (%) of 74 (majority vote), 72.9** (mean) [confidence interval 69-76.7], accuracy (%) of 68 (majority vote), 67.7** (mean) [65.9-69.5], sensitivity (%) of 60 (majority vote), 60.6** (mean) [52.7-68.6], specificity (%) of 74 (majority vote), 73.3** (mean) [67.8-78.9], PPV (%) of 65 (majority vote), 64.6** (mean) [62-67.1], and NPV (%) of 70 (majority vote), 70** (mean) [67.1-72.9] (*p<0.05, **p<0.005).
Conclusion: Radiomics-based machine learning may assist clinicians in predicting clinical outcome of patients with lumbar spine arthrodesis based on post-operative X-rays, thus modifying physical rehabilitation and therapeutic strategies accordingly.
Limitations: Retrospective study.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by Local Ethics Committee (RETRORAD protocol)

Author Block: J-B. Pialat¹, D. Gicquel¹, A-K. Golla², C. Bürger², C. Lorenz², M. Villien¹, S. Gouttard¹, A. Vlachomitrou¹, T. Klinder²; ¹Lyon/FR, ²Hamburg/DE
Purpose: AI algorithms which detect vertebral fractures generate limited classifications which only identify vertebral body fracture. We propose a thoracolumbar and sacral spine fracture detection algorithm able to identify individual fracture locations in both the vertebral body and the posterior arch . It segments the entire spine, extracts spine-aligned sub volumes and detects spinal fractures using a convolutional neural network.
Methods or Background: 195 CT scans from polytraumatized patients were collected in a single-center retrospective clinical study. Dataset was split into training (n=145) and validation (n=50) sets. Accuracy for identification of injury location within the body was assessed in the validation set using a Free Response ROC (FROC) curve and performance at the vertebral body level was measured using a Receiver Operating Characteristic (ROC) curve. A subsequent test set including 173 patients ( fractured N=109, non fractured N= 64) was analyzed with the same algorithm. Performance was assessed similarly using FROC curve.
Results or Findings: The algorithm detected 87.3% of the 775 spinal fracture locations of the validation set using a false positive threshold of 5 per case. It detected 92.4% of the fractured vertebrae. 249 false positives were detected, most of which were easily rejected upon review by radiologists. 26 false negatives were found, most of which were transverse process fractures. There were 7 vertebral body fractures; all were single endplate stable fractures. In the test set, the algorithm detected 88.6% of the 255 fractures using a false positive threshold of 5 per case.
Conclusion: We have developed and validated a deep learning algorithm which determines location of fractures in the whole vertebra with reasonable accuracy.
Limitations: This as to be tested prospectively in routine emergency condition to assess the gain in time / sensitivity
Funding for this study: Collaborative study between Hospices Civils de Lyon and Philips using GOPI research fundings
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by local ethics commitee

Author Block: G. Del Gaudio¹, G. Vuurberg², M. Dalmau-Pastor³, G. Kerkhoffs⁴, M. Maas⁴; ¹Rome/IT, ²Weesp/NL, ³Barcelona/ES, ⁴Amsterdam/NL
Purpose: In the literature, there is much conflicting data regarding the anatomy of the anterior tibiofibular ligament (ATiFL), even in studies with anatomical specimens. Therefore, this study aims to reassess the anatomy of this ligament using MRI with a high-resolution isotropic 3D-PDw sequence.
Methods or Background: From February to May 2024, 72 MRI scans (3T) of the ankle were performed at Amsterdam UMC. The inclusion criterion was patients over 16 years of age. The exclusion criteria were: absence of a 3D-PDw scan, ATiFL trauma or surgery, congenital anomalies, metallic or movement artifacts. The 43 3D-PDw valid scans, allowed for aligning the planes along the individual bundles of the ligament.
Results or Findings: The high spatial resolution (0,23mm) of 3DPDw allowed the identification of three bundles: superior, intermediate and inferior. Regarding dimensions the superior is the thickest and widest (mean 2.68x9.28mm) and the inferior the longest (mean 15.45mm). Regarding orientation (axial from the fibula to the tibia) the superior and inferior have a transverse orientation, while the intermediate is oriented backward. Regarding the shape, they are fanned in 97.7%, 71.7%, and 25.6% respectively, while they are band-like in the remaining cases. We did not identify any anatomical variance regarding the number of bundles.
Conclusion: The use of volumetric isotropic sequences as the 3D-PDw, can be a very useful tool for the anatomical study of ligamentous structures in the absence of available anatomical specimens. Understanding the exact anatomy of this structure is crucial for managing both acute and chronic traumatic pathologies (impingement, overuse), especially in young patients and athletes.
Limitations: Sample size and lack of anatomical specimen comparison.
Funding for this study: No funding.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study received a waiver by the ethical committee according to local rules and regulations.

Author Block: Y. Song, H. Xu, R. Xu, K. Xu; Chengdu/CN
Purpose: To evaluate the disease progression in patients with Duchenne muscular dystrophy (DMD) by using multi-modal quantitative magnetic resonance imaging (qMRI), and comparing the responsiveness of these imaging indicators with the clinical function scales.
Methods or Background: 130 DMD patients were enrolled and underwent MRI examination of hip muscles to determine fat fraction (FF) and longitudinal relaxation time (T1). All participants returned for follow-up at an average of 12 months. According to the baseline North Star Ambulatory Assessment (NSAA) score, all patients were divided into three subgroups: mild (76-100 score), moderate (51-75 score) and severe (0-50 score) functional decline. Standardized response mean (SRM) was used as the responsiveness to the disease progression, and the responsiveness of qMRI and clinical function scales to the disease progression in different DMD stages was compared. SRM>0.8 is considered as a high response to disease progression.
Results or Findings: The overall SRM of MRI biomarkers is higher than that of the clinical function scales. For mild group, FF of adductors and abductors have higher responsiveness, with SRM of 0.816 and 1.043, respectively. For moderate group, FF of all muscle groups have a high responsiveness, and the SRM are between 1.004 and 1.606. For severe group, T1 of abductors and FF of all muscle groups have high responsiveness, and SRM are between 0.867 and 1.633. However, the SRM of the clinical function scales for patients with different disease stages are all less than 0.8.
Conclusion: The sensitivity of MRI biomarkers to DMD disease progression is higher than that of clinical function scales, especially the FF of gluteal muscles is more sensitive to disease progression, and the sensitivity indicators are different in different disease stages.
Limitations: This study didn‘t discuss whether patients received steroid therapy.
Funding for this study: National Natural Science Foundation of China (82271981)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: ChiCTR1800018340

Author Block: T. D. Lerch, T. Kaim, K. Ziebarth, M. K. Meier, J. D. Busch; Bern/CH
Purpose: Slipped capital femoral epiphyses (SCFE) is a common pediatric hip disease with the risk of osteoarthritis and impingement deformities, and 3D models could be useful for patient-specific analysis. Therefore, magnetic resonance imaging (MRI) bone segmentation was investigated.
Methods or Background: A retrospective IRB-approved study involving 23 symptomatic pediatric patients (23 hips) with SCFE was performed. All patients underwent preoperative hip MR with pelvic axial high-resolution images (T1 VIBE DIXON images). Slice thickness was 1.2 mm. Mean age was 12 ± 2 years. All patients underwent surgical treatment.
Manual and automatic MRI-based bone segmentation was compared.
automatic bone segmentation was performed by machine learning algorithm, a previously used and validated convolutional neural network trained for adult pelvis bone segmentaiton was adapted to pelvis of children.
Results or Findings: Manual MRI-based bone segmentation was feasible (all patients, 100%, duration 4-5 hours per case).
Dice coefficient was calculated to assess differences between manual and automatic bone segmentation,
Dice coefficient was 82% for the pelvis and 88% for proximal femur.
Precision was 80% for the pelvis and 94% for proximal femur.
Conclusion: MRI-based 3D models were feasible for SCFE patients. Three-dimensional models could be useful for SCFE patients for preoperative 3D printing and deformity analysis.
This could aid for patient-specific diagnosis, treatment decisions, and preoperative planning.
MRI-based 3D models are radiation-free and could be used instead of CT-based 3D models in the future for computer-assisted 3D simulation of surgery.
Limitations: MRI is expensive and access is limited
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB approval was obtained

Author Block: E. Nikolova¹, J. Kroschke¹, C. Obermüller¹, F. Zecca², K. Pawlus¹, T. Rauer¹, F. Ensle¹; ¹Zurich/CH, ²Cagliari/IT
Purpose: To assess whether a single isotropic 3D proton-density-weighted fat-saturated (PDFS) sequence could replace a standard 2D multisequence MRI protocol for comprehensive examination of the knee using deep learning reconstruction (DLR).
Methods or Background: In this retrospective study, 95 consecutive patients > 18 years without history of prior knee surgery undergoing MRI knee examination at the same 1.5 Tesla scanner between May 2023 and July 2024 were included. Standard MRI protocol with DLR consisted of a 3D PDFS sequence and five 2D fast-spin-echo sequences in various orientations. Two radiologists separately evaluated the 3D sequence in all three planes and the 2D sequences, assessing pathologies of bone, cartilage, menisci and ligaments for all joint compartments, and overall image quality, diagnostic confidence and artifacts. Wilcoxon signed-rank test was used to compare Likert scale gradings, McNemar’s test for binary grades. Interreader agreement was assessed with Cohen’s kappa.
Results or Findings: There was no significant difference between protocols regarding assessment of medial(MC) and lateral compartment(LC) meniscus, (MC) and patellofemoral(PF) cartilage, medial and lateral collateral ligament, anterior and posterior cruciate ligament, MC and PF bone marrow edema(BME), and fractures in all compartments (p>0.05). Significant differences were shown in assessment of LC cartilage (p=0.002) and LC BME (p=0.04). Image quality and artifacts did not demonstrate significant differences. Diagnostic confidence was significantly higher for the 2D protocol(p=0.023). Interreader agreement overall was substantial for the 3D-PDFS(k=0.67) and 2D protocol (k=0.66).
Conclusion: Our results suggests comparable performance between a single 3D-PDFS and a multisequence 2D protocol using DLR for comprehensive assessment of knee structures, except for LC cartilage and BME. With DLR-powered image enhancement, 3D-PDFS might be able to partly replace 2D sequences for time-efficient knee MRI in the future.
Limitations: Retrospective study design. No arthroscopic reference standard.
Funding for this study: This research received no financial support.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Not applicable

Author Block: I. Ristow¹, S. Zhang², C. Riedel¹, A. Lenz¹, M. Krause¹, G. Adam¹, P. Bannas¹, F. O. Henes¹, L. Well¹; ¹Hamburg/DE, ²Best/NL
Purpose: To evaluate the feasibility and diagnostic performance of a 3D FRACTURE (fast field echo resembling a CT using restricted echo-spacing) MRI sequence for the detection and classification of proximal tibial fractures compared with CT.
Methods or Background: We retrospectively included 126 patients (85 male; 39.6±14.5 years) from two centers following acute knee injury. Patients underwent knee MRI at 3T including FRACTURE-MRI. Additional CT was performed in patients with tibial fractures (32.5%; n=41) as the reference standard for fracture classification. Two radiologists independently evaluated FRACTURE-MRI for the presence of fractures and classified them according to AO/OTA, Schatzker, and the 10-segment classification. Diagnostic performance of FRACTURE-MRI was assessed using crosstabulations. Inter-reader agreement was estimated using Krippendorff’s alpha. Image quality was graded on a five-point scale (5=excellent; 1=inadequate definition of fracture lines and fracture displacement) and assessed using estimated marginal means.
Results or Findings: Fractures were detected by FRACTURE-MRI with a sensitivity of 91.5% (83.2–96.5%) and a specificity of 97.1% (93.3–99.0%). Regarding fracture classification, diagnostic performances were slightly lower, with the 10-segment classification yielding the best sensitivity of 85.7% (81.4–89.3%) and specificity of 97.4% (96.6–98.0%), and the Schatzker classification yielding the lowest sensitivity of 78.2% (67.4–86.8%) and specificity of 97.7% (94.1–99.4%). Inter-reader agreement across the whole cohort was excellent (Krippendorff’s alpha 0.89–0.96) and when considering only patients with fractures, good to acceptable (0.48–0.91). Image quality was rated good (estimated marginal mean 4.3 (4.1–4.4)).
Conclusion: FRACTURE-MRI is feasible at 3T enabling accurate delineation of fracture lines for precise diagnosis and classification of proximal tibial fractures.
Limitations: Future studies need to address in a comparative intra-individual setting whether the diagnostic performance of FRACTURE-MRI is better or equivalent to other CT-like bone imaging techniques, such as UTE/ZTE, GRE, or SWI, for fracture detection.
Funding for this study: N/A
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The retrospective study was approved by the local institutional review board (Ärztekammer Hamburg).

Author Block: A. Ferrero, J. Thorne, A. O. El Sadaney, K. Rajendran, C. Mccollough, F. Baffour; Rochester, MN/US
Purpose: Pathologies affecting bone health impact both mineral density (vBMD) and morphometric characteristics (thickness (Th) and spacing (Sp)) of trabecular (Tb) and cortical (Ct) bone. This work introduces a semi-automated software, qBone, which quantifies bone morphometry from in vivo CT scans of the extremities and the vertebral spine.
Methods or Background: protocols were optimized for extremity and spine exams using a commercial photon-counting-detector (PCD) CT system. A dedicated CNN algorithm was trained to reduce image noise of the spine CT exams while maintaining high resolution details. Adaptive segmentation algorithms automatically delineated Ct and Tb compartments allowing quantification of Th, Sp and vBMD for each. To validate the software’s accuracy, 10 cadaveric wrists were scanned on HRpQCT and PCD-CT. A 3D-printed bone model (Ct.Th=2mm, Tb.Th=0.3mm and Tb.Sp=0.75mm) was used to assess the CNN denoising performance across different patient sizes. Finally, qBone was applied in vivo to multiple prospective cohorts for wrist and spine.
Results or Findings: optimized PCD-CT protocols for wrist (70kV, 12mGy, <0.1mSv) and spine (120kV, 40mGy, 8mSv) yielded <0.15mm in-plane resolution. Validation with cadaveric wrists and the 3D-printed bone model demonstrated excellent agreement in Ct.Th and Tb.Sp metrics. CNN denoising significantly improved trabecular morphometry accuracy in the spine for small and medium patient sizes. In vivo measurements (wrist, N=50; spine, N=14) for each metric (Tb.Th=0.3-0.45mm, Tb.Sp=0.6-1.05mm, Ct.Th=0.5-1.58mm, Ct.vBMD=450-600mg/cm3) were consistent with literature values.
Conclusion: qBone facilitates semi-automated quantification of bone morphometry from high resolution CT data, providing a comprehensive assessment of bone health in vivo beyond traditional mineral density.
Limitations: Comparisons with microCT are needed to validate metrics for vertebral bones. Additionally, qBone does not leverage spectral information to estimate vBMD in the spine.
Funding for this study: NIH R21ar084126-01a1
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB 23_005308, PI: Baffour