Study of the Impact of Image Quality and Quantity on Neural Network Training Outcome

The article examines the impact of image quality and quantity on the training results of a convolutional neural network, using dental radiographs as an example of real-world medical application.The role of input data in the processes of classification and pathology recognition is analyzed, and the key factors influencing the accuracy of deep learning models in medical imaging are identified.Neural network performance was modeled using images of varying quality (blur, added noise, reduced resolution) and different quantities of training examples.It was demonstrated that degradation of visual data quality significantly reduces network training efficiency even with a large number of examples, whereas high-quality images support consistently high model accuracy even with limited data volume.The main patterns of classification errors were determined, revealing that models most frequently misclassify classes with similar visual characteristics.The study employed the approach of analyzing results through confusion matrices and accuracy change graphs during training and validation.Based on the conducted research, the necessity of ensuring high data quality for medical artificial intelligence models was substantiated, and directions for improving neural network architectures through the integration of attention mechanisms and the use of data augmentation techniques were proposed.The results can be utilized for further development of automated diagnostic systems for dental diseases and for creating more flexible neural network solutions capable of operating under conditions of unstable data quality.