Search

Objectives: To develop and validate machine learning models to distinguish between benign and malignant bone lesions and compare the performance to radiologists., Methods: In 880 patients (age 33.1 ± 19.4 years, 395 women) diagnosed with malignant (n = 213, 24.2%) or benign (n = 667, 75.8%) primary bone tumors, preoperative radiographs were obtained, and the diagnosis was established using histopathology. Data was split 70%/15%/15% for training, validation, and internal testing. Additionally, 96 patients from another institution were obtained for external testing. Machine learning models were developed and validated using radiomic features and demographic information. The performance of each model was evaluated on the test sets for accuracy, area under the curve (AUC) from receiver operating characteristics, sensitivity, and specificity. For comparison, the external test set was evaluated by two radiology residents and two radiologists who specialized in musculoskeletal tumor imaging., Results: The best machine learning model was based on an artificial neural network (ANN) combining both radiomic and demographic information achieving 80% and 75% accuracy at 75% and 90% sensitivity with 0.79 and 0.90 AUC on the internal and external test set, respectively. In comparison, the radiology residents achieved 71% and 65% accuracy at 61% and 35% sensitivity while the radiologists specialized in musculoskeletal tumor imaging achieved an 84% and 83% accuracy at 90% and 81% sensitivity, respectively., Conclusions: An ANN combining radiomic features and demographic information showed the best performance in distinguishing between benign and malignant bone lesions. The model showed lower accuracy compared to specialized radiologists, while accuracy was higher or similar compared to residents., Key Points: • The developed machine learning model could differentiate benign from malignant bone tumors using radiography with an AUC of 0.90 on the external test set. • Machine learning models that used radiomic features or demographic information alone performed worse than those that used both radiomic features and demographic information as input, highlighting the importance of building comprehensive machine learning models. • An artificial neural network that combined both radiomic and demographic information achieved the best performance and its performance was compared to radiology readers on an external test set., (© 2022. The Author(s).)

Background An artificial intelligence model that assesses primary bone tumors on radiographs may assist in the diagnostic workflow. Purpose To develop a multitask deep learning (DL) model for simultaneous bounding box placement, segmentation, and classification of primary bone tumors on radiographs. Materials and Methods This retrospective study analyzed bone tumors on radiographs acquired prior to treatment and obtained from patient data from January 2000 to June 2020. Benign or malignant bone tumors were diagnosed in all patients by using the histopathologic findings as the reference standard. By using split-sample validation, 70% of the patients were assigned to the training set, 15% were assigned to the validation set, and 15% were assigned to the test set. The final performance was evaluated on an external test set by using geographic validation, with accuracy, sensitivity, specificity, and 95% CIs being used for classification, the intersection over union (IoU) being used for bounding box placements, and the Dice score being used for segmentations. Results Radiographs from 934 patients (mean age, 33 years ± 19 [standard deviation]; 419 women) were evaluated in the internal data set, which included 667 benign bone tumors and 267 malignant bone tumors. Six hundred fifty-four patients were in the training set, 140 were in the validation set, and 140 were in the test set. One hundred eleven patients were in the external test set. The multitask DL model achieved 80.2% (89 of 111; 95% CI: 72.8, 87.6) accuracy, 62.9% (22 of 35; 95% CI: 47, 79) sensitivity, and 88.2% (67 of 76; CI: 81, 96) specificity in the classification of bone tumors as malignant or benign. The model achieved an IoU of 0.52 ± 0.34 for bounding box placements and a mean Dice score of 0.60 ± 0.37 for segmentations. The model accuracy was higher than that of two radiologic residents (71.2% and 64.9%; P = .002 and P < .001, respectively) and was comparable with that of two musculoskeletal fellowship-trained radiologists (83.8% and 82.9%; P = .13 and P = .25, respectively) in classifying a tumor as malignant or benign. Conclusion The developed multitask deep learning model allowed for accurate and simultaneous bounding box placement, segmentation, and classification of primary bone tumors on radiographs. © RSNA, 2021 Online supplemental material is available for this article. See also the editorial by Carrino in this issue.