Source-linked AI summary
Deep learning predicts total knee replacement from magnetic resonance images
Aniket A. Tolpadi, Jinhee J. Lee, Valentina Pedoia, Sharmila Majumdar
TL;DR
The paper addresses prediction of TKR risk, particularly before OA reaches its most advanced stages. It combines MRI with clinical and demographic information and analyzes occlusion maps; the integrated model reports AUC 0.834±0.036 overall and 0.943±0.029 for patients without OA.
Problem
The authors seek to predict which patients will undergo TKR, especially at less advanced OA stages where nonsurgical treatments remain options.
Source: Introduction · p. 1Method
A deep-learning pipeline combines knee MRI predictions with clinical and demographic information, stratifies logistic-regression ensembles by OA severity, and uses occlusion maps to examine imaging biomarkers.
Source: Introduction · p. 2, Pipeline architecture · p. 4, Discussion · p. 11Results
The integrated MRI pipeline achieved AUC 0.834±0.036 overall and AUC 0.943±0.029 for patients without OA.
Source: Discussion · pp. 10–11Takeaways & Limitations
The authors conclude that risk prediction could support non-invasive treatment for patients with no or moderate OA, while the biomarker analysis identifies tissues associated with TKR risk.
Source: Discussion · p. 12Only 12 test-set cases without radiographic OA underwent TKR within five years, limiting the available evidence for this subgroup.
Source: Discussion · p. 12
Abstract
from arXiv · showhide
Knee Osteoarthritis (OA) is a common musculoskeletal disorder in the United States. When diagnosed at early stages, lifestyle interventions such as exercise and weight loss can slow OA progression, but at later stages, only an invasive option is available: total knee replacement (TKR). Though a generally successful procedure, only 2/3 of patients who undergo the procedure report their knees feeling ''normal'' post-operation, and complications can arise that require revision. This necessitates a model to identify a population at higher risk of TKR, particularly at less advanced stages of OA, such that appropriate treatments can be implemented that slow OA progression and delay TKR. Here, we present a deep learning pipeline that leverages MRI images and clinical and demographic information to predict TKR with AUC $0.834 \pm 0.036$ (p < 0.05). Most notably, the pipeline predicts TKR with AUC $0.943 \pm 0.057$ (p < 0.05) for patients without OA. Furthermore, we develop occlusion maps for case-control pairs in test data and compare regions used by the model in both, thereby identifying TKR imaging biomarkers. As such, this work takes strides towards a pipeline with clinical utility, and the biomarkers identified further our understanding of OA progression and eventual TKR onset.
Introduction
TKR prediction could help guide treatment before OA reaches its late stages, when TKR is the only option described. Existing image-based prediction evidence is limited, and this study combines knee imaging with clinical and demographic information while evaluating performance by OA severity.
Introduction
- At late OA stages (KL = 4), TKR is the only option described, while earlier stages can be managed with nonsurgical treatments.The motivation is to identify higher-risk patients so appropriate treatments can be pursued, potentially prolonging knee health and delaying TKR.Source: Introduction · p. 1
- Earlier TKR imaging-biomarker studies primarily examined cartilage volume loss, subchondral bone defects, and bone marrow lesions.The authors argue that identifying additional biomarkers could improve understanding of OA and TKR and guide treatment strategies.Source: Introduction · p. 2
- Prior TKR prediction models using images were limited, and earlier work had not reported sensitive and specific prediction at early OA stages in severity-imbalanced datasets.The paper notes that overall AUC can be inflated when severe OA cases are easier to predict and more prevalent among eventual TKR patients.Source: Introduction · p. 2
- Deep learning’s automated feature extraction and the availability of large cohorts such as OAI support its feasibility for complex TKR prediction.The authors situate their study after deep learning had been applied to knee OA classification and progression prediction.Source: Introduction · p. 2
- The study combines knee images with clinical and demographic information to predict TKR and evaluates performance across OA severity.It also uses occlusion maps to identify imaging biomarkers for eventual TKR onset.Source: Introduction · p. 2
Novelty
The paper identifies three novel aspects: 3D DenseNet MRI-based TKR prediction, evaluation stratified by OA severity, and comprehensive case-control identification of imaging biomarkers.
Novelty
- The study is the first to apply a 3-dimensional DenseNet CNN to predict TKR from MRI.Source: Novelty · p. 2
- The model is evaluated separately by OA severity, an analysis the authors say previous studies had not reported.Source: Novelty · p. 2
- The study reports a comprehensive case-control analysis to identify imaging biomarkers for TKR.The authors frame this analysis as aimed at improving interpretability and clinical utility.Source: Novelty · p. 2
Materials and Methods
The study uses OAI knee images and non-imaging variables to train an OA-stratified pipeline for predicting TKR within five years. It also compares occlusion hotspots in matched cases and controls to examine imaging biomarkers.
Materials and Methods
- MRI volumes were center-cropped to 120×320×320, while radiographs were cropped around the knee; both image types were normalized.MRI pixel values were rounded to 14 possible values after the authors observed better validation performance than with unrounded values.Source: Data · p. 3
- A DenseNet-121 was pretrained for OA prediction and fine-tuned for TKR, with its output combined with 27 non-imaging variables.The resulting 28 variables entered one of three logistic-regression ensembles for no, moderate, or severe OA.Source: Data · p. 3, Pipeline architecture · p. 4, Training · p. 5
- The dataset includes 35,482 entries from 4,790 unique patients, with patient-level separation across training, validation, and test sets.The OAI cohort provided 2D radiographs and 3D DESS MRI images over 10 years; cases were first TKRs within five years.Source: Data · pp. 2–3, Table 2 · p. 4
- Logistic regression was selected to integrate imaging and non-imaging predictions, with separate ensembles optimized by OA severity and model outputs averaged for final predictions.Model comparisons used McNemar’s test for correct predictions and bootstrap AUC comparisons for sensitivity-specificity performance.Source: Integration of Imaging and Non-Imaging… · p. 5, Statistical Analysis · p. 5
Imaging biomarker identification
- For 124 integrated-MRI true positives, controls were sampled from test-set true negatives with matching OA-status distributions and KOOS pain scores.Occlusion hotspots were pixels in the top 5% of output changes; tissue hotspot frequencies were compared between cases and controls to identify significant regions.Source: Imaging biomarker identification · pp. 5–6
Results
The results report MRI pipeline performance overall and by OA severity, compare MRI with X-ray, and identify tissue hotspots associated with TKR risk.
Results
- OA-pretrained models provided useful information, but fine-tuning and integration of non-imaging variables were needed for the desired TKR prediction performance.The radiograph pretrain had poor sensitivity without OA and poor specificity in moderate and severe OA; the MRI pretrain had remaining sensitivity and specificity gaps.Source: OA pretrain utility in TKR prediction · p. 6
MRI pipeline optimization and performance
- MRI AUC was 0.834±0.036 overall and 0.943±0.029 for the integrated model in patients without OA.The image-only MRI AUC was 0.886±0.020 overall and 0.897±0.039 without OA.Source: MRI pipeline optimization and… · p. 6, Figure 2 · p. 8
- The integrated MRI pipeline outperformed X-ray by AUC for patients without OA and with severe OA, but not moderate OA or overall.The corresponding comparison p values were 3.04×10−2, 4.57×10−2, 9.55×10−1, and 9.94×10−1, respectively.Source: Comparison of MRI and radiograph… · p. 7, Figure 3 · p. 9
- Integrated X-ray achieved higher accuracy, while integrated MRI had higher sensitivity across OA stages, markedly so without OA.The authors note that this sensitivity-accuracy trade-off favors MRI when identifying patients at risk of eventual TKR is paramount.Source: Comparison of MRI and radiograph… · p. 7, Table 4 · p. 7
- Without OA, adding non-imaging variables increased X-ray AUC from 0.514±0.087 to 0.799±0.055 and MRI AUC from 0.897±0.039 to 0.943±0.029.The authors report that non-imaging variables seemed to add more critical information to X-ray than to MRI.Source: Comparison of MRI and radiograph… · p. 8
- Hotspots in the medial patellar retinaculum, gastrocnemius tendon, and plantaris muscle were associated with higher TKR risk; tibiofemoral cartilage, bone, menisci, and ACL hotspots with lower risk.The authors interpret these contrasting associations as suggesting TKR onset is more nuanced than OA progression, with prediction relying less on tissues around the tibiofemoral joint.Source: Biomarker identification and analysis · pp. 8–9
Discussion
The discussion reports strong MRI-based TKR prediction, especially for patients without radiographic OA, and identifies candidate imaging biomarkers. The authors also note limits in the OAI sample and occlusion-map resolution.
Discussion
- The integrated pipeline reported accuracy of 78.5±0.134%, sensitivity of 81.8±0.643%, and specificity of 78.4±0.138%.The authors report that MRI outperformed X-ray for patients without OA and with severe OA, and suggest MRI may contain information about non-imaging variables.Source: Discussion · p. 10
- The MRI image-only model achieved AUC 0.886±0.020, compared with 0.86±0.01 for Wang et al.’s DESS or TSE image-only models; the integrated MRI model achieved AUC 0.834±0.036.Wang et al.’s combined image-and-non-imaging model achieved AUC 0.88±0.02.Source: Discussion · p. 10
- For patients without OA, the MRI pipeline achieved an AUC of 0.943±0.029; sensitivity and specificity were 92.2±1.68% and 82.4±0.173% at no OA, respectively.At moderate OA, sensitivity and specificity were 78.9±0.974% and 74.7±0.228%; the integrated model was optimized for Youden’s index within OA classifications rather than overall AUC.Source: Discussion · p. 11
- Occlusion hotspots in medial and lateral cartilage and bone, meniscal regions, and the ACL were associated with reduced TKR risk, while three other tissues were associated with increased risk.The increased-risk hotspots were the medial patellar retinaculum, gastrocnemius tendon, and plantaris muscle; Table 5 summarizes hotspot-frequency comparisons using odds ratios and significance tests.Source: Discussion · pp. 11–12, Table 5 · p. 11
- The OAI sample limits generalizability and statistical power: only 66 patients without radiographic OA underwent TKR within five years, including 12 test-set cases.The study also compressed MRI pixels to 14 possible values for performance, and used occlusion-map voxels of 12×32×32 with stride 12 because smaller settings were computationally infeasible.Source: Discussion · p. 12
- The authors conclude that prediction may help identify patients for nonsurgical treatment to delay TKR, while calling for improved image-feature integration, alternate architectures, and faster predictions.They propose further study of the plantaris muscle and gastrocnemius tendon because their roles in OA progression and TKR are not well characterized.Source: Discussion · p. 12
Supplementary Figures
Figure 5 shows sample DESS MRI slices alongside compressed versions created by rounding pixel values after normalization.
Supplementary Figures
- Figure 5 pairs sample DESS MRI slices with versions compressed by rounding normalized pixel values.The figure illustrates the effect of the pixel-value compression step; the caption does not report a performance comparison.Source: Figure 5 · p. 15
Supplementary Tables
The supplementary tables document candidate non-imaging variables and stratified hotspot frequencies for integrated MRI predictions.
Supplementary Tables
- Table 6 lists literature-identified non-imaging variables correlated with OA progression or eventual TKR that were considered for supplementing image-based predictions.Variables were included as potential additions when present in the OAI database.Source: Table 6 · p. 16
- Table 7 reports percentages of selected tissue hotspots among 124 true positives detected by the integrated MRI pipeline, stratified by OA severity.Source: Table 7 · p. 17
- Table 8 reports percentages of selected tissue hotspots among 124 true-negative controls detected by the integrated MRI pipeline, stratified by OA severity.Source: Table 8 · p. 18