A brain tumor computer-aided diagnosis method with automatic lesion segmentation and ensemble decision strategy.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Gliomas and brain metastases are common brain malignancies with distinct prognoses.
• Accurate tumor type diagnosis is crucial for effective treatment and patient outcomes.
• Traditional radiomics pipelines lack integrated segmentation and classification, hindering diagnostic efficiency.

Purpose of the Study:

• To develop an improved computer-aided diagnosis method for gliomas and brain metastases.
• To integrate automatic lesion segmentation and ensemble decision strategies into a radiomics pipeline.
• To enhance the accuracy and efficiency of brain tumor diagnosis using multi-center MRI data.

Main Methods:

• Utilized preoperative MRI scans (T1-CE and T2-flair) from 1,022 glioma and 775 brain metastasis patients across three hospitals.
• Developed two segmentation models for automatic tumor segmentation, followed by radiomics feature extraction.
• Implemented machine learning classifiers, a weight soft voting model, and an ensemble decision strategy based on prior knowledge.

Main Results:

• The proposed pipeline achieved an accuracy (ACC) of 0.8950 and an area under the receiver operating characteristic curve (AUC) of 0.9585.
• These results surpassed the traditional radiomics pipeline's performance (ACC: 0.8850, AUC: 0.9450).
• The automated segmentation significantly improved diagnostic metrics.

Conclusions:

• The developed model accurately classifies gliomas and brain metastases using MRI radiomics.
• The novel pipeline demonstrates high generalizability and interpretability for brain tumor diagnosis.
• This approach shows significant potential for improving clinical diagnosis of brain malignancies.