¹Unfallkrankenhaus Berlin, Berlin, Germany
²Berliner Hochschule für Technik, Berlin, Germany
³Klinik für Orthopädie und Unfallchirurgie am Brüderklinikum Julia Lanz, Mannheim, Germany

Text

Introduction: Incidental findings occur in up to 20% of conventional radiographs [1]. The complex morphology of bone tumors, inflammatory changes, normal variants, or metastases can lead to unnecessary diagnostic escalation (MRI, contrast-enhanced imaging, scintigraphy) [2]. This causes costs, radiation exposure, biopsy of a do-not-touch lesion [3], and in the worst case, contamination of the access pathway in malignant bone tumors. Previous Machine Learning (ML) models primarily aim at radiological diagnosis [4], not at selecting the treatment pathway. Particularly in facilities not specialized in tumor orthopedics, the recurring question is: “Continue diagnostics or refer to a specialized center?” Using explainable ML methods (“explainable AI”), a demonstrator for a Clinical Decision Support System (CDSS) will be developed to address this question.

Methods: Sarcomas (including malignant fibrous histiocytoma), suspicious osteochondromas, adamantinomas, chordomas, and Langerhans cell histiocytosis were recommended for referral. Conventional radiographs with demographic data and histologically confirmed diagnoses were collected from radiological case collections and literature. These were annotated according to standardized features. Multiple ML models (Decision Trees, Naive Bayes) were trained and evaluated against a dataset withheld during training (stratified 4-fold cross-validation).????

Results: 1031 cases (male: 606, female: 425) were included in the dataset. These contained 49 different diagnoses (ranging from intraosseous lipoma and osteomyelitis to parosteal osteosarcoma). The Naive Bayes model correctly assigned an unknown lesion to the appropriate pathway in 80.27% of validation cases, while the Decision Tree model achieved 79.66%. Age, ill-defined tumor margins, and cortical breakthrough were the most discriminative features. In contrast, the models correctly identified the specific diagnosis (from 49 possibilities) in only 38% of cases.

Conclusion: Complex radiographic morphology, typical age peaks, and a high number of possible diagnoses create a high-dimensional problem that is difficult for humans but relatively simple for ML models to map. The presented results show that by modifying the question from “computer makes diagnosis” to “which pathway is appropriate”, a significant improvement in accuracy can be achieved. In a hypothetical scenario of night duty in a primary and standard care provider, this decision is more relevant than the definitive diagnosis. Therefore, goals and questions for CDSS should always be developed in interdisciplinary dialogue between medicine and computer science, considering the deployment scenarios.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.

Literatur

[1] Blackburn CW, Richardson SM, DeVita RR, Dong O, Faraji N, Wurtz LD, et al. What Is the Prevalence of Clinically Important Findings Among Incidentally Found Osseous Lesions? Clin Orthop Relat Res. 2023 Oct;481(10):1993.
[2] Miller BJ, Avedian RS, Rajani R, Leddy L, White JR, Cummings J, et al. What Is the Use of Imaging Before Referral to an Orthopaedic Oncologist? A Prospective, Multicenter Investigation. Clin Orthop. 2015 Mar;473(3):868–74.
[3] Vanel D, Ruggieri P, Ferrari S, Picci P, Gambarotti M, Staals E, et al. The incidental skeletal lesion: ignore or explore? Cancer Imaging Off Publ Int Cancer Imaging Soc. 2009 Oct 2;9 Spec No A(Special issue A):S38-43.
[4] Ong W, Zhu L, Tan YL, Teo EC, Tan JH, Kumar N, et al. Application of Machine Learning for Differentiating Bone Malignancy on Imaging: A Systematic Review. Cancers. 2023 Mar 18;15(6):1837.