Background/Objectives: Large language models (LLMs) are revolutionizing the world and the field of medicine while constantly improving themselves. With recent advancements in image interpretation, evaluating the reasoning capabilities of these models and benchmarking their performance on brain MRI tasks has become crucial, as they may be utilized—albeit off-label—for patient care by both neurosurgeons and non-neurosurgeons. Methods: ChatGPT-4o, Grok, and Gemini were presented with 35,711 slices of brain MRI, including various pathologies and normal MRIs. Models were asked to identify the MRI sequence and determine the presence of pathology. Their individual performances were measured and compared with one another. Results: GPT refused to answer 28.02% of the slices despite three attempts, whereas Grok and Gemini provided responses on the first attempt for every slice. Gemini achieved 74.54% pathology prediction and 46.38% sequence prediction accuracy. GPT-4o achieved 74.33% pathology prediction and 85.98% sequence prediction accuracy for questions that it had answered (53.50% and 61.67% in total, respectively). Grok achieved 65.64% pathology prediction and 66.23% sequence prediction accuracy. Conclusions: The image interpretation capabilities of the investigated LLMs are limited for now and require further refinement before competing with specifically trained and fine-tuned dedicated applications. Amongst them, Gemini outperforms the others in pathology prediction while Grok outperforms others in sequence prediction. These limitations should be kept in mind if use during patient care is planned.

Early diagnosis of brain tumors is extremely important, and shortening the interval between the acquisition of MRI images and reporting of the results is critical for patients. In the diagnosis of brain tumors, CT and MRI are some of the core diagnostic techniques used today. Our main goal is to reduce the workload of radiologists by developing a neural network that segments MRI images of the brain so we propose a multi-path segmentation algorithm based on U-Net architecture that uses residual extended skip blocks. Our proposed model is trained and tested with Gazi Brains 2020 Dataset. We evaluated the results using the dice similarity coefficient and compared the results with other segmentation algorithms and saw that our proposed model has comparatively better results. Our proposed model is using T1-Weighted, T2-Weighted, and Flair MRI images together as inputs, whereas other segmentation models, are using T2-Weighted or Flair MRI images as input. Implementation of the model and trained models are available at https://github.com/batuhansozer/brain-segmentation-with-novel-multi-path-model