Chronic osteomyelitis diagnosis using MRI: optimising radiomics models through multi-level region of interest extensions.

PURPOSE: Chronic osteomyelitis is a persistent infectious bone disease characterised by high recurrence, in which surgical debridement plays a crucial role. Conventional debridement mainly depends on surgeons' clinical experience and visual interpretation of imaging findings, which are inherently subjective and may not accurately define the boundary between infected and healthy tissue. Radiomics demonstrates substantial potential in this domain; however, it faces critical challenges, including region of interest (ROI) selection and spatial mapping. To enhance the diagnostic accuracy of radiomics in osteomyelitis and support preoperative evaluation, this study introduces an innovative multi-level ROI expansion framework. This approach aims to enable comprehensive and precise lesion characterisation as well as to provide a quantitative reference for determining the extent of pathological involvement.

PATIENTS AND

METHODS

Clinical and imaging data from 102 patients with suspected chronic osteomyelitis of the long bones were retrospectively reviewed. The following three ROI models were constructed: an initial ROI model, a 5-mm expansion ROI model and a 10-mm expansion ROI model. Each model underwent radiographic feature analysis and was analysed using four machine learning algorithms: K-nearest neighbour, support vector machine, random forest and logistic regression. Diagnostic performance was assessed using receiver operating characteristic curve analysis, clinical impact curve evaluation and performance metrics including area under the curve (AUC), sensitivity, specificity and accuracy.

RESULTS

Logistic regression algorithm consistently outperformed the other algorithms across all ROI expansion ranges. Compared with the 0-mm and 10-mm ranges, all models demonstrated superior performance within the 5-mm feature range. The DeLong test revealed significant differences among the logistic regression submodels (p < 0.05), with the 5-mm expansion model achieving the highest AUC.

CONCLUSION

The proposed multi-level ROI expansion framework enhances the radiomics-based diagnosis of chronic osteomyelitis. By comparing diagnostic performance across expanded ROI models, we demonstrated that the combination of a 5-mm expanded ROI and logistic regression yielded the highest diagnostic efficacy. These findings indicate that the perilesional region contains critical pathological information, thereby providing theoretical support for the clinical strategy of moderate debridement expansion. Furthermore, the 10-mm expanded ROI model demonstrated distinct advantages in detecting more extensive tissue alterations. Collectively these two ROI expansion strategies offer surgeons a more comprehensive lesion assessment and data-driven insights, thereby facilitating precise surgical planning.