Table 1 Comparison of deep learning and impedance-based detection in melanoma detection
From: Wearable battery-free chip-less patch for bioimpedance measurement of cutaneous lesions
Challenge | Limitations of Deep Learning (Optical Imaging) | Advantages of Impedance Based Detection |
|---|---|---|
Skin tone representation | Mostly performs well on fair skin tones, underperforms on darker skin due to biased image datasets. | Works equally well across all skin tones as it measures electrical properties, not visual appearance. |
Class imbalance | Struggles with class imbalance, often misclassifying visually similar lesions (e.g., melanoma vs. benign). | Focuses on bioimpedance differences, improving accuracy in distinguishing visually similar lesions54. |
Generalization to clinical settings | May not generalize well to diverse clinical environments due to dataset bias and variability in imaging. | Works across varied clinical settings, unaffected by lighting or skin tone. |
Small lesion detection | Small lesions lack clear visual features, reducing detection sensitivity. | Can detect the changes even for small lesions that may not be visually distinctive. |
Clinical applicability and acceptability | Often criticized as a “black box,” making clinical trust and adoption difficult. | Provides quantifiable biophysical metrics such as impedance (resistance-capacitance), aiding clinical interpretation. |
Privacy and data concerns | Requires large volumes of patient image data, raising privacy and storage concerns. | Typically generates nonvisual, numerical data, reducing privacy risks and data storage burden. |
