Purpose:
Manual rating of Cerebral microbleeds (CMBs) is time-consuming and inconsistent. Since the presence and number of CMBs have become a potential diagnostic and prognostic biomarker of stroke, an automatic identification method is required. We proposed a computer aided diagnosis (CAD) system for the detection of the CMBs on the magnetic resonance (MR) images automatically.
Methods:
Eighty-one patients were recruited in this study. CMBs on the MR T2* weighted images were manually rated according to the Microbleed Anatomic Rating Scale (MARS) criteria. Our automated method consisted of two steps: i) Pre-processing: After skull stripping, isolated islands of points were removed while holes were restored to avoid over segmentation. Local threshold segmentation was applied for the initial candidate selection. ii) Identification model: Seven features were extracted from each candidate: area, roundness, intensity, average of the boundary, contrast, shape-intensity and location-mark (according to the probability density templates calculated from the location information of the CMBs). For further identification of each candidate, Random Forest (RF) model was used to distinguish CMBs from the mimics.
Results:
A total of 337 CMBs in the 81 patients were studied. Comparing with the counting from the experienced doctors, high sensitivity of 92% (310/337) was achieved after pre-processing. The RF model eliminated most of the false-positives while maintaining a reliable sensitivity of 94% (291/310) and specificity of 96% (4272/4450). The area under the Receiver operating characteristic curve was 0.98 ± 0.02 for the detection model. In summary, this CAD system had an overall sensitivity of 86% (291/337) and specificity of 96% (4272/4450), producing only 2.2 false-positives per subject.
Conclusion:
This presented strategy is technically effective. The results indicate that it has the potential to be used for clinical detection of CMBs.
Development of a Computer-aided Diagnosis System to Distinguish between Benign and Malignant Mammary Tumors in Dynamic Magnetic Resonance Images: Automatic Detection of the Position with the Strongest Washout Effect in the Tumor
