Differentiating childhood benign and malignant bone lesions in radiological examination using artificial intelligence

Abstract

Background: Primary bone tumors in children and adolescents are rare, therefore posing a significant diagnostic challenge. Malignant lesions are critical diagnoses that require intensive treatment and can be fatal. Early detection and reliable assessment of the lesion’s nature are of great clinical importance. The aim of this thesis was to create a dataset of radiographic images of pediatric benign and malignant bone lesions and to use it to evaluate the ability of EfficientNet models (B0 – B7) in classifying lesion dignity. Patients and Methods: The study included patients aged 0 to 19 years who underwent radiographical examination between 01.01.2004 and 30.05.2024. Diagnoses were either confirmed histopathologically or, if biopsy was not necessary on unavailable, made based on characteristic radiological appearance. To minimize confounding factors, only pre-interventional images were included, and those taken after biopsy, surgery, chemotherapy, or radiotherapy were excluded. The final dataset consisted of 800 X-ray images from 228 individual patients. These images were used to train and test EfficientNet variants B0 to B7. Evaluation metrics included accuracy, recall, precision, and F1-score. Additionally, Precision-Recall (PR) and Receiver Operating Characteristic (ROC) curves were created and the area under the curve (AUC) was calculated. Results: The patients’ ages ranged from 0.10 to 17.60 years. The dataset included 396 images (38.25%) from female and 464 images (61.75%) from male patients. Of the total, 707 lesions (88.37%) were benign and 93 (11.63%) were malignant. All EfficientNet models demonstrated very high recall (0.979 – 1.000) and good precision (0.909 – 0.914) in detecting benign bone tumors. The highest F1-score for benign lesions was achieved by EffiecientNet-B2 (0.954). The performance on malignant lesions was significantly lower. Recall ranged between 0.237 to 0.301, with precision values varying between 0.651 and 1.000. The corresponding F1-scores were low (0.379 – 0.424). Overall, EfficientNet-B2 showed the best performance across all cases, with an accuracy of 0.915 and an F1-score of 0.689. The highest PR-AUC was achieved by B3 (0.624), closely followed by B1, B2, and B5. Discussion: EfficientNet models are well suited for detecting benign pediatric bone tumors, as demonstrated by their high recall in this particular group. However, reliability for malignant lesions was inadequate, with an average recall of only 27%. The models correctly identified malignant lesions only when classification certainty was high, resulting in low false-positive rate but a high false-negative rate. The main reason for these results is likely due to the significant class imbalance in the dataset (88% benign vs. 12% malignant), which greatly limits model generalizability. For reliable application in clinical practice, the model performance for malignant cases must be improved. This can be achieved by expanding the dataset, for instance through synthetic data generation or data augmentation.

Datum der Bewilligung	2025
Originalsprache	Englisch
Gradverleihende Hochschule	Medizinische Universität Graz
Betreuer/-in	Sebastian Tschauner (Betreuer*in) & Anja Dutschke (Mitbetreuer*in)