EXPLORING THE POTENTIAL OF NEURAL NETWORKS TO PREDICT METASTATIC FEMUR STRENGTH FROM 2D PROJECTIONS Alexander Synek (1), Emir Benca (2), Dieter H. Pahr (1,3) 1. TU Wien, Austria; 2. Medical University of Vienna, Austria 3. Karl Landsteiner University of Health Sciences, Austria Introduction Estimating the fracture risk of femora with metastatic lesions represents a major challenge in clinical practice, resulting in overtreatment of many patients [1]. Finite element (FE) models have proven to be a useful tool for fracture risk assessment and are on the verge of clinical use [2]. However, they still require 3D imaging and substantial computational resources. Predicting the femoral strength from 2D projections, such as plain radiographs, would enhance clinical availability, allow continuous monitoring and could, therefore, further improve the patients’ quality of life. The goal of this study was to explore the potential of artificial neural networks (NN) for fast, quantitative strength predictions of metastatic femora from 2D projections. In a first step, we tested if a NN can predict the strength of a single femur with various different metastatic lesions. Methods To conduct a systematic exploration of the potential of NN for femoral strength prediction, an artificial dataset was created (Fig. 1): First, a 3D computed tomography (CT) image of an intact proximal femur of a previous study [3] was selected. Second, artificial metastatic lesions were randomly inserted into the image, leading to 2000 variations of different lesions within the same femur. Third, non-linear FE models were used to predict the strength (Fmax) in stance configuration for each bone. The FE modelling workflow was based on [4], but adapted for efficiency and calibrated using experimental data of 68 human femur specimens [3,4] (Fmax predicted with R²=0.89). Finally, anterior-posterior 2D projections were created for each of the bones. Thus, a dataset was obtained consisting of 2000 2D projections and the corresponding FE-predicted femoral strengths. The dataset was split into a training (n=1800) and validation dataset (n=200). The training dataset was used to train a custom convolutional NN that takes a single 2D projection as an input and predicts the femoral strength. The validation was performed using the validation dataset by predicting the femoral strength 1) using the original 2D projections and 2) using the 2D projections with small random rotations and translations to test the robustness of the NN predictions (Fig. 2). Goodness of fit (R²) and mean absolute percentage error (MAPE) were used to evaluate the performance of the NN. Results The NN predicted the femoral strength with R²=0.95 and MAPE of 8.4% using the original 2D projections, and R²=0.93 and MAPE of 9.6% using the randomly rotated and translated 2D projections (Fig. 2). Discussion This study showed that a convolutional NN can quantitatively predict the femoral strength from 2D projections with good precision and accuracy for a single femur with different artificial metastatic lesions. Random rotations and translations of the 2D projections only slightly influenced the results. In the next steps, the dataset shall be expanded to include multiple femora, more realistic representations of the metastatic lesions, and finally CT scans of patients with real metastatic lesions. References 1. Benca et al, Bone Rep., 5:51-56 2016. 2. Eggermont et al, Cancers, 14:5904, 2022. 3. Benca et al, Sci. Rep., 9:10305, 2019. 4. Dall’ara et al, Bone, 1:27-38, 2013