Radiomics based on multiple machine learning methods for diagnosing early bone metastases not visible on CT images

Area of Science:

• Medical Imaging
• Radiology
• Oncology

Background:

• Early detection of bone metastases is crucial for patient outcomes.
• Computed tomography (CT) may not detect microscopic early bone metastases.
• Radiomics analysis of CT images offers potential for enhanced detection.

Purpose of the Study:

• To evaluate radiomics features from CT combined with machine learning for identifying microscopic early bone metastases.
• To determine the optimal machine learning approach for this task.
• To use [99mTc]-methylene diphosphate (MDP) single photon emission computed tomography (SPECT) as a reference standard.

Main Methods:

• Retrospective study of 63 patients with early bone metastasis.
• Registration of SPECT-defined volumes of interest (VOIs) onto CT images.
• Extraction of 944 radiomics features from 126 VOIs (63 metastasis, 63 normal bone).
• Development and evaluation of 20 machine learning models using 5 feature selection and 4 classification methods.

Main Results:

• Most models achieved an area under the receiver operating characteristic curve (AUC) > 0.70.
• K-Nearest Neighbors (KNN) classifier showed superior performance.
• The XGBoost feature selection with KNN classifier achieved the highest AUCs: 0.989 (training) and 0.975 (testing).

Conclusions:

• Radiomics and machine learning can identify early bone metastases not visible on CT.
• The KNN classifier is the optimal method for this application.