Evaluation of the effectiveness of a deep learning model based on EfficientNetB3 for differential diagnosis of Alzheimer's disease stages

The purpose of the research is evaluation of the effectiveness of the modified EfficientNetB3 architecture based on transfer deep learning and early stopping methods in medical decision support systems for differential diagnosis of Alzheimer's disease stages. Methods . To conduct experimental studies, a training dataset was generated, normalized, and augmented. A modified EfficientNetB3 neural network architecture was implemented using transfer learning and early stopping methods in Python. The neural network model was trained. Results . The classification performance of the trained neural network model was assessed using the Recall, Precision, Specificity, F1-score, and AUC-ROC metrics. Analysis of these metrics revealed that the results achieved by the modified EfficientNetB3 architecture are characterized by significant asymmetry, indicating the highly specialized nature of this model. On the one hand, the model proved to be an effective tool for diagnosing moderate dementia, demonstrating the highest possible AUC value. On the other hand, classification performance for the remaining classes was significantly lower (AUC values for the "No Dementia," "Very Mild Dementia," and "Mild Dementia" classes were 0,87, 0,86, and 0,95, respectively). Conclusion . Based on the results of the analysis, it can be concluded that the primary practical value of this modification of the EfficientNetB3 architecture lies in its use in heterogeneous ensembles or cascaded diagnostic systems for verifying a specific stage of Alzheimer's disease – moderate dementia – in order to improve the overall system efficiency. This points to the potential for further research in the area of creating highly specialized architectures capable of solving specific subproblems with high accuracy, surpassing general-purpose but less focused approaches.