Comparison Between Convolutional Neural Networks and Human Pattern Recognition in Neuroradiological Images

Pattern recognition in medical images is a classical problem in computational analysis, with direct applications in supporting clinical diagnosis. In neurology, the interpretation of magnetic resonance imaging (MRI) exams requires a high level of visual discrimination due to the complexity of brain structures and the variability of pathologies affecting the central nervous system.
In this work, an approach based on convolutional neural networks is presented for the identification of neuroradiological patterns and its comparison with the performance of human observers. A neural network model was developed and trained using previously preprocessed images, and its classification capability was evaluated through traditional machine learning metrics such as accuracy, precision, and sensitivity.
In parallel, a set of technical questionnaires containing the same images was administered to groups of human participants, considered in this study as biological neural networks. In the initial stage of the analysis, the results indicated that there was no statistically significant correlation between the participants’ level of knowledge and their performance in pattern recognition tasks. In contrast, the neural network achieved higher accuracy in all comparisons performed.
In the second stage of the research, an automatic image selection algorithm was implemented to identify the most representative images in the dataset, considering parameters related to visual quality and diagnostic informativeness. This new set of images was then used in a second application of the questionnaires to the same group of participants.
The results showed an improvement in human performance and a reduction in the accuracy gap between participants and the computational model. Statistical analyses were conducted using absolute and relative frequencies, chi-square tests with a significance level of 5%, and metrics of absolute and relative variation between performances.
The findings indicate that artificial intelligence techniques can contribute not only to the advancement of automated medical image analysis but also to the development of quantitative methodologies to support education and training in neurology.