RESUMEN
Artificial intelligence (AI) in medical care can raise diagnosis accuracy and improve its uniformity. This study developed a diagnostic imaging system for chronic wounds that can be used in medically underpopulated areas. The image identification algorithm searches for patterns and makes decisions based on information obtained from pixels rather than images. Images of 50 patients with pressure sores treated at Kobe University Hospital were examined. The algorithm determined the presence of necrosis with a significant difference (p = 3.39 × 10-5). A threshold value was created with a luminance difference of 50 for the group with necrosis of 5% or more black pixels. In the no-necrosis group with less than 5% black pixels, the threshold value was created with a brightness difference of 100. The "shallow wounds" were distributed below 100, whereas the "deep wounds" were distributed above 100. When the algorithm was applied to 24 images of 23 new cases, there was 100% agreement between the specialist and the algorithm regarding the presence of necrotic tissue and wound depth evaluation. The algorithm identifies the necrotic tissue and wound depth without requiring a large amount of data, making it suitable for application to future AI diagnosis systems for chronic wounds.
RESUMEN
Deep learning is used to classify data into several groups based on nonlinear curved surfaces. In this paper, we focus on the theoretical analysis of deep learning using the rectified linear unit (ReLU) activation function. Because layers approximate a nonlinear curved surface, increasing the number of layers improves the approximation accuracy of the curved surface. While neurons perform a layer-by-layer approximation of the most appropriate hyperplanes, increasing their number cannot improve the results obtained via canonical correlation analysis (CCA). These results illustrate the functions of layers and neurons in deep learning with ReLU.
