gms | German Medical Science

Objectives: The KIDS (“Künstliche Intelligenz zur Diagnostik von Sarkomen bei Kindern”) project aims, among others goals, to address a significant challenge in pediatric oncology: the early detection of benign, intermediate and malignant bone tumours in children with x-ray imaging. While Artificial Intelligence (AI) has shown proficiency in differentiating between two classes of tumours [1], a critical gap remains in initial tumour detection, particularly for non-oncology-trained professionals and general practitioners. Pediatric tumours are often incidentally discovered, underscoring the need for an AI system capable of identifying suspicious areas in children's X-ray images.

Methods: This retrospective study utilized X-ray data from a diverse cohort of pediatric patients from our local musculoskeletal tumour database. The dataset comprised 817 images (567 pathological and 250 healthy) from 511 patients, including the following entities: aneurysmal bone cyst, chondroblastoma, chondrosarcoma, enchondroma, Ewing’s sarcoma, fibrous dysplasia, giant cell tumour, non-ossifying fibroma, osteochondroma, osteosarcoma. We employed the ResNet18 architecture for classification, supported by robust cross-validation techniques and data augmentation strategies. Our methodology focused on enhancing the AI system’s ability to generalize across various clinical scenarios.

Results and conclusion: The AI model demonstrated exceptional performance with an accuracy of 96.39%, a sensitivity of 96.0%, and a specificity of 96.0%. The variance in cross-validation splits was 0.05, 0.10, and 0.13, respectively, indicating highly stable results across different test sets. These metrics reflect the model's reliability and potential effectiveness in clinical settings.

Current AI applications in orthopedic oncology are progressing yet remain insufficiently robust for widespread clinical use. However, our findings underscore the significant potential of AI tools in aiding both young professionals and general practitioners. The early detection of bone tumours in children, facilitated by AI, can profoundly impact patient outcomes. This underscores the urgent need for further development and integration of AI in pediatric oncology diagnostics. Additionally, future advancements should focus on multimodal approaches [2] that incorporate not only X-ray data but also MRI and, crucially, clinical data. Integrating these diverse data sources will enhance the accuracy and applicability of AI in diagnosing and managing pediatric bone tumours, offering a more holistic and effective approach to patient care.

Figure 1 [Fig. 1]