Introduction Osteoporosis, characterized by reduced bone mass and microarchitectural deterioration [1], is typically identified by declines in bone mineral density (BMD) scores measured through dual-energy X-ray absorptiometry (DXA) [2]. It's associated with an increased risk of fragility fractures, including hip and vertebral fractures, which carry significant morbidity and a 25–35% mortality rate at 1 year post-hip fracture [3]. Osteoporosis is often considered a "silent" disease due to the absence of obvious symptoms. Spine MRIs are a routine part of daily MRI scans. Using daily lumbar MRI evaluations as an opportunistic osteoporosis screening tool can aid in early detection. Recent studies have reported reduced osteogenesis and increased adipogenesis in aging bone [4], increased adipose tissue and reduced bone formation in osteoporosis [5]. The state-of-the-art MRI technique, IDEAL-IQ, offers precise fat and water mapping capabilities [6]. Therefore, IDEAL-IQ can be applied opportunistically for screening lumbar vertebral osteoporosis. Radiomics-based machine learning (ML) models are an emerging field in clinical medicine, extensively used for precise disease detection, diagnosis, prediction, and prognosis [7].The Tree-based Pipeline Optimization Tool (TPOT), an automated machine learning (AutoML) algorithm, autonomously designs and optimizes ML pipelines for specific problem domains without human intervention [8]. This study aimed to develop a TPOP-radiomics classification model for discriminating between normal and osteoporosis using fat and water MR images. Methods Image Data: We enrolled 65 patients with osteoporosis and 64 normal subjects of corresponding age to the patients, as identified by BMD measures using DeXA. Fat, water, fat fraction, and water fraction images were obtained through quantitative MRI using the IDEAL-IQ method on sagittal images. A total of 10 to 14 slices covered the lumbar spine from the T12 vertebra to the sacrum. The region of interest (ROI) of lumbar vertebra (L1~L4) was segmented and then reviewed with slice-by-slice by a senior radiologist. Figure 1 illustrates the TPOP-radiomics model workflow. Feature Extraction and Selection: We employed Pyradiomics (version 1.3.0) to automatically extract radiomic features. Feature selection involved several steps: First, we retained Radiomics features with significant differences using Student's t-test (p < 0.05). Second, we selected features by combining the results of four optimal screening methods, which included ANOVA, L1 regularized LASSO regression, mutual information, and linear support vector machines (SVM). Finally, the selected features underwent normalization for scaling. TPOT: In this study, we applied TPOT (version 0.12.1), an AutoML tool for Python that leverages scikit-learn to search for the best model using its default configuration, which includes all data operators and machine learning classification models. The TPOT search was set to run for 20 generations and 30 population size sets with five-fold cross-validation. Additionally, we performed 10 random repetitions for the dataset in which the training and test data were split at a ratio of 70% for training and 30% for testing. Three datasets were used to compare the performance of AutoML-radiomics model in discriminating between normal and osteoporosis. These datasets (Figure 2) were generated based on a model with the largest lumbar area slice (2D), a model with whole lumbar and the largest lumbar area slice (2D+3D), and a model with reconstructed 2D map (Recon2D) from whole lumbar volume including max intensity projection (MIP), min intensity projection (mIP), and average. Performance evaluation: In comparing all the models, various performance evaluations were employed, including sensitivity, specificity, and the area under the curve (AUC). Results Table 1 displays the average performance parameters for the best model across the three datasets. Figure 3 presents box plots for the AUC, sensitivity, and specificity of the three models across 10 random repetitions. Figure 4 displays the ROC curves for these models. Among all the models studied, Recon2D demonstrated the best performance with mean sensitivity of 0.745 and mean specificity of 0.758 in distinguishing between normal and osteoporosis. Table 2 displays the performance results of Recon2D across 10 random repetitions. The types of radiomic features among these major features are presented in Table 3. Discussion and Conclusion In this study, we employed a TPOP-radiomics-based model using fat and water MRI for automatic ML model selection to distinguish between normal and osteoporosis. While previous research has shown the effectiveness of fat fraction MRI and clinical data for this discrimination [9], our study focused solely on fat and water MRI, potentially enhancing the scope of osteoporosis opportunistic screening. The Recon2D model exhibited the best performance, which could be attributed to BMD measurements being 2D projections in DeXA. In summary, our study demonstrates that the TPOP-radiomics-based model enables opportunistic osteoporosis screening using fat and water MRI.

Acknowledgements

Supported by Taiwan National Science and Technology Council under grants 111-2314-B-035 -001 -MY3, and by the Taichung Armed Forces General Hospital (Grant no.: 107A42 and 108A16).