Rapid and reliable diagnosis of cerebral stroke is a vital necessity, and among them, ischemic stroke is the most difficult to recognize on MRI images. Increasing the efficiency of stroke diagnosis is associated with the transition to incre asingly “heavy” AI tools working in 3D mode, processing MRI images not pixel-by-pixel, but voxel-by-voxel and using complex multifactor information processing algorithms. The implementation of such products requires large computing resources, which are often unavailable outside large medical centers. In addition, existing explainable AI tools identify the affected area very roughly and generally, which reduces the doctor’s confidence in the diagnostic result. On the other hand, one can use AI models operat ing in 2D mode - weaker, but faster and less demanding on computing resources. The article uses a Siamese neural network as a base model. To improve classification efficiency, a model pretrain on “light” synthetic data based on Perlin noise is proposed. To objectify the choice of mode, the article uses the apparatus of topological data analysis, namely, changes in persistent entropy and Renyi entropy of data embeddings on fully connected layers of a neural network. It is experimentally confirmed that using a 2D model, when trained on slices with maximum lesion visibility, produces ROC-AUC values no worse than using a full-scale 3D SOTA models, while allowing the clinician to selectively evaluate the individual slices he or she selects. It is experimentally c onfirmed that simple model that contains only locally useful features can support neural network training to a level comparable to much more complex and resource-intensive generative model.