(P 171) THREE-DIMENSIONAL DEEP LEARNING MODEL FOR PREDICTING PROXIMAL FEMORAL PATHOLOGIC FRACTURE USING OPPORTUNISTIC ABDOMINOPELVIC CT IN PATIENTS WITH ADVANCED CANCER: A PRELIMINARY REPORT

Objective: Proactive surgical intervention for proximal femoral pathologic fractures in patients with advanced cancer is crucial, and opportunistic use of follow-up abdominopelvic CT scans can help predict them. This study aimed to develop and validate a 3D deep learning model to predict the fractures.

Methods: A total of 2,933 abdominopelvic CT scans of 540 patients were collected from four tertiary academic hospitals. Since patients with bilateral proximal femoral metastatic lesions contributed two different cases from a single CT scan, 3,621 cases were finally included. There were 1,956 and 1,665 scans from females and males, respectively, with a mean age of 63 ± 11 years. The scans were labeled into fracture (F) and non-fracture (N) groups based on fracture occurrence within 3 months post-CT. Images were clipped to a bone window, normalized, and converted to 3D volumes using anterior-inferior-iliac-spine-based segmentation. From this landmark, 32 consecutive axial slices were extracted to create three-dimensional input volumes. The training dataset, from a single institution, included 1,884 N- and 12 F-labelled scans, and the latter was augmented 20-fold through affine transformations. Internal validation used 472 N- and 12 augmented F-labelled scans. An external test set comprised 1,241 scans from the other institutions: 1,214 N- and 27 F-labelled scans. Performance of four 3D deep learning models (ResNet-18, ResNet-34, DenseNet-121, and Swin Transformer) was evaluated.

Results: DenseNet-121 apparently showed the highest external test performance, (accuracy 0.9315, sensitivity 0.8483, specificity 0.9703, precision 0.9315, F1 score 0.9315). ResNet-34 followed closely (accuracy 0.9242, sensitivity 0.7743, specificity 0.9942, precision 0.9242, F1 score 0.9242), showing better performance than ResNet-18 (accuracy 0.8782, sensitivity 0.8201, specificity 0.9053, precision 0.8782, F1 score 0.8782). Swin Transformer showed lower performance (accuracy 0.5244, sensitivity 0.3686, specificity 0.5972, precision 0.5244, F1 score 0.5244).

Conclusion: The three-dimensional deep learning model alone demonstrated promising performance. Evaluating how model-assisted predictions influence clinical decision-making is essential to establish the model’s clinical utility. To improve the model’s robustness and generalizability, an extensive dataset will be derived from the Clinical Data Warehouse.