A fuzzy evaluation approach to determine superiority of deep learning network system in terms of recognition capability: case study of lung cancer imaging

Artificial intelligence (AI) assists in decision-making across various fields and industries. Diverse market needs have prompted the rapid evolution of AI learning algorithms. Deep learning networks (DLNs) process classification problems associated with perceptrons; this approach has become mainstream in the current AI era. To compare the classification and recognition performances of the designed DLN systems, most studies have applied confusion matrices as assessment tools and further computed accuracy, sensitivity, and specificity indices for judgment and analysis. However, the values of these indices change with the degree of learning achieved by the network system each time it is trained. Thus, using a single index value or mean value to determine recognition capabilities may lead to misjudgment. In view of this, we used accuracy to define a recognition performance index (RPI) $$I_{ACC}$$ I ACC . Considering the unavoidable uncertainty in $$I_{ACC}$$ I ACC , we further propose a triangular fuzzy number (TFN) for $$I_{ACC}$$ I ACC . This is applied to develop a fuzzy test model for $$I_{ACC}$$ I ACC to aid researchers in evaluating superiority among the designed DLN systems in terms of recognition capabilities. To demonstrate the applicability of the proposed approach, we implemented it on a LeNet-5 convolutional neural network system optimized using the Taguchi method for tomographic images of lung cancer provided by the 2015 International Society for Optics and Photonics (SPIE).