Eurospine, Vienna, Avusturya, 2 - 04 Ekim 2024, cilt.4, ss.3, (Özet Bildiri)
Brain and Spine 4 (2024) 103227
Radiomics-Powered Radiographic Image Analysis for Enhanced
Mechanical Complications Prediction and Surgical Planning in Adult
Spine Deformity
F. Pellise 1, S. Haddad 2, S. Nú~ nez-Pereira 2, C. Yilgor 3, M. Barcheni 2, A. Pupak 2
,
M. Ramirez 2, J. Pizones 4, A. Alanay 3, I. Obeid 5, F. Kleinstück 6, F. Galbusera 6
,
O. Sagarra 7, E.S.S.G. Essg 2
.
1 Spine Unit Orthopaedic Department, Vall d'Hebron
University Hospital, Barcelona, Spain2 Spine Research Unit, Vall d'Hebron
University Hospital, Barcelona, Spain3 Spine Unit Orthopaedic Department,
Acibadem Mehmet Ali Aydinlar University,_
Istanbul, Türkiye4 Spine Unit
Orthopaedic Department, La Paz University Hospital, Madrid, Spain5 Spine Unit
Orthopaedic Department, H^ opital Pellegrin Bordeaux, Bordeaux,
France6 Department of Spine Surgery, Schulthess Klinik, Zürich, Switzerland7 Data
Research, Dribia Data Research, Barcelona, Spain
Introduction: Radiomics, a technique employing machine learning to extract
quantitative features from processed radiographic images, holds promise for
improving clinical prediction models. It offers the potential to comprehensively
characterize spinal shape and alignment. We Hypothezise that Processed image
(PrIm) algorithms outperform traditional radiographic measurements (TRM) in
predicting postoperative mechanical complications (MC) in adult spinal defor-
mity (ASD).
Materials and Methods: Retrospective study analyzing prospective preop, 6-
week and 2-year follow-up data of surgical ASD patients. Processed full-spine
standing radiographic images were analyzed using an automatic vertebral and
pelvic centroid generation algorithm to map postero-anterior and lateral spinal
shape. Distances and angles between each vertebral centroid and pelvis were
automatically obtained. Machine learning models were constructed (Catboost),
combining non-radiographic variables (Non-R: demographic, PROMS, surgical),
TRM (Cobb, C7SVL, PI, PT, SS, LL, TK, GT, SVA, GAP score) and PrIm features.
AUC-ROC, sensitivity, speci city, and Brier score (0¼ perfectly calibrated /
1¼poor) were used to evaluate prediction accuracy. SHAP values were employed
to assess variable contributions and address over tting/noise.
Results: 690 patients (81% female, 52 19 years, 9.7 3.9 levels, 18.6% 3CO,
43.5% pelvic xation, 24.3% MC) were analyzed. The PrIm + Non-R model
outperformed the present“Gold Standard” model (TRM + Non-R): AUC-ROC
0.75 vs 0.71 (p¼0.009), accuracy 0.72 vs 0.62 (p<0.001), speci city 0.79 vs 0.60
(p<0.001), sensitivity 0.52 vs 0.70 (p<0.001), and Brier score 0.17 vs 0.21
(p<0.001). Adding TRM to PrIm+Non-R model did not improve model estimates
(Fig). SHAP adjusted models summed 35 variables and revealed PrIm's superior
predictive importance, contributing 65.7% to the model compared to Non-R
(Surgical factors 16.1%, PROMS 11.3% and demographics 6.9%). Personalized
SHAP decision plots identi ed critical vertebral centroids associated with MC
risk, both globally and individually.
Conclusion: Radiomics, powered by full-spine processed radiographic images,
enable the most accurate predictive models for mechanical complications in ASD.
This novel approach offers clinicians a powerful and time-ef cient tool for
personalized surgical planning.
AUC-ROC versus the number of features sorted by importance.