A White-Box Module for Medical Image Multi-Classification via Bias Dendrite Net

The deep neural network has promoted the design of excellent multi-classification methods for medical images. Owing to attractive advantages: white-box attribute, strong logical expression capability, controllable precision for better generalization capability, and lower computational complexity, the dendrite net has found wide applications in classification, regression, and system identification. However, ignorance of the translation invariance constrains the flexibility of its application in medical image classification. In this study, we propose the bias dendrite net (BDD Net) for medical image classification by introducing linear and Hadamard product operations to establish the logical relationships among inputs. Five different BDD Net modules are designed and optimization standards for BDD Net modules are defined. The experimental results on the dataset MedMNIST V2 show our BDD Net (Module II) outperforms the other four modules, so BDD Net (Module II) is the final choice of our BDD Net. Experiments are conducted to compare our BDD Net (Module II) with the dendrite net and black-box models on MedMNIST V2. Our BDD Net (Module II) obtains the average accuracy of 88.5%, the average area under curve (AUC) of 96.5%, the average precision of 85.9%, the average recall of 88.5%, and the average ${F}1$ of 86.9%, which are better than the results obtained by the baselines. We also validate the applicability of our BDD Net (Module II) on the chest X-ray dataset, the experimental results show that our method maintains a good classification performance.