Integrated Boundary-Overlap-Size Metric for Local Assessment of Deep Learning Methods in Medical Microwave Imaging

Deep learning has been a game-changer in enhancing the speed and accuracy of medical microwave imaging in detecting abnormal lesions. Nonetheless, the challenge lies in establishing a universal objective metric to assess the reliability of these methods. Current evaluation practices often rely on a single geometric metric, which presents inherent constraints. Consequently, the evaluations of results generated by deep learning methods may not always reflect clinicians’ insights and judgments. To overcome this, a local assessment metric incorporating the following three geometric dimensions is proposed: the overlap between the detected anomaly and the actual lesion, the proximity of their boundaries, and the proportionality of the lesion sizes determined by the algorithm versus the actual lesion. This approach to evaluation ensures that the resulting metric's score is in line with professional medical diagnostics. The presented results on head imaging using five deep learning algorithms confirm the significant advantages of the proposed metric, validating its effectiveness in providing objective evaluation of various algorithms in medical electromagnetic imaging. This objective metric is poised to guide future algorithm development to ensure a reliable assessment of their capability in abnormality detection and diagnosis.