Workflow

Split and Labelled fMRI Data for Deep Learning

Figure 1 shows the process of segmenting and labeling fMRI data. We collected more than 1,000 participants’ data in the seven aspects: emotional, gambling, language, sports, relationship, social and working memory (WM) functional magnetic resonance as the model training data set, performed 3T and MRI on these data, then, standardized the preprocessed fMRI data to reduce training errors. Each task input sample in the data set is a continuous 3D sequence. After excluding non-brain regions and adding the time dimension, we obtained a total of more than 34,000 fMRI 4D data, of which 60% were used as training data and 40% were used as verification data and test data.

Figure 1 Split and Labelled fMRI Data for Deep Learning

Deep Neural Network Classification and Visualization

Figure 2 and Figure 3 show our CNN network classification and classification results visualization process. The 3DCNN classification network in the brain decoding model is composed of some 3D convolution layers and fully connected layers. The convolution kernel in the convolution layer generates a time description and extracts high-dimensional features for each voxel of the input data. The fully connected layer uses the softmax function and contains 64 feature channels and 7 categories (respectively corresponding to 7 features of the data set). Finally, we use instructive back-propagation to visualize the deep network, create a corresponding 3D model by drawing each voxel with an absolute maximum in the time domain, and then map it to the fsaverage surface.

Figure 2 Deep Neural Network Classification

Figure 3 Visualization Results

Advantages

• Accurate
  The deep neural network classification model in our brain decoding model is data-driven and trained from a large number of fMRI data. It has more stable accuracy than manual analysis. The average accuracy of 7 tasks is 94%: emotion (94%), gambling (83%), language (97%), sports (97%), relationship (89%), social (96%) and WM (91%), under the ROC curve The average area is 0.996.
• Efficient
  The deep neural network classification model in our brain decoding model runs on a high-performance GPU server and works 24/7 without interruption, which can accurately and quickly complete fMRI data analysis and visual analysis results. In addition, our brain decoding model also provides a series of data statistics Tools to help researchers complete their work more efficiently and with higher quality.
• Intuitive
  The brain decoding model in our artificial intelligence system has powerful data visualization tools, presenting the results of processing and analysis in a more intuitive and easy-to-read way, which brings a pleasant user experience to researchers while improving work efficiency.

CD ComputaBio has formed a team of experts excellent in imaging science and clinical domain knowledge, providing AI-driven solutions for Brain Decoding according to your detailed requirements.