Machine Learning Matches Human Performance at Segmenting the Human Visual Cortex.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Neuroimaging

Background:

• Accurate localization of visual areas on the human cortical surface is crucial for neuroscience research.
• Current methods, including human raters, face challenges with inter-rater reliability, time commitment, and data collection requirements.
• Existing automated methods lack the accuracy achieved by human experts.

Purpose of the Study:

• To train convolutional neural network (CNN) models to predict visual area boundaries and iso-eccentric regions (V1, V2, V3).
• To compare the accuracy of CNNs using functional versus anatomical data against human raters.
• To investigate the relationship between anatomical structure and functional organization in visual cortex.

Main Methods:

• Training CNN models on the Human Connectome Project and NYU Retinotopy datasets.
• Utilizing both functional and anatomical MRI data for model training and prediction.
• Comparing CNN performance against human expert raters and existing automated methods.

Main Results:

• CNNs trained on functional data achieved accuracy comparable to human raters.
• CNNs trained solely on anatomical data showed lower but superior accuracy compared to existing automated methods.
• Eccentricity mapping showed less correlation with anatomical structure than polar angle mapping, indicating inter-subject variability.
• Approximately 75% of V1, V2, and V3 could be accurately mapped using only structural MRI data.

Conclusions:

• CNNs offer a powerful, accurate, and efficient tool for mapping visual brain areas, approaching human expert reliability.
• Machine learning techniques, particularly CNNs, are poised to become integral to future neuroscience research for brain mapping.
• The study reveals a tighter-than-expected coupling between cortical structure and function in early visual areas.