Research on MedMNIST medical image classification and recognition method using improved Bayesian neural network

Rapid lesion identification in medical images is crucial for early warning and treatment of conditions. Aiming at the problem that images vary and it is challenging to establish a robust uncertainty estimation model, a MedMNIST medical image classification and recognition method based on Bayesian theory is proposed. This method uses ResNet18 as the baseline network, and some units use Bayesian convolutional layers and Bayesian fully connected layers to efficiently extract discriminative features, suppress background interference, and improve model training quality; the ResNet18_BNN framework is introduced to combine the weights in the model with The convolution kernel parameters are transformed from point estimation to probability distribution, and a comprehensive cost and risk loss function is designed for optimization; By comparing the performance of models such as ResNet50, MobileNetV2, MobileViT-XXS and ConvNeXt-T on the MedMNIST data set, quantitative and qualitative analysis, The generalization ability of the model was verified. Experimental results show that the improved model has an accuracy of 0.895 in some data sets, which is 0.052 higher than ResNet18, 0.038 higher than ResNet50, and 0.077, 0.061, and 0.063 higher than the lightweight SOTA detection algorithms MobileNet, MoBileViT, and ConvNeXt respectively. It shows that the algorithm has high accuracy, effectively improves the recognition accuracy, and provides ideas for model selection and performance optimization for subsequent work.