CT image analysis of intracranial hemorrhage by deep learning
1. Learning training process:
We used 25,000 published cases of head CT as teaching data.
The label of the teaching data includes the presence or absence of intracranial hemorrhage. In addition, five types of labels, (1) epidural, (2) intraparenchymal, (3) intraventricular, (4) subarachnoid, and (5) subdural, are also added.
As a learning model, we used EfficientNet published by Google AI.
The images below are multiple examples of teaching data for subdural bleeding.
2. Inference process:
After learning, we used the new head CT images and let the post-learning AI engine software make inferences. The results were evaluated for the presence or absence of bleeding and the classification of bleeding types. As a result of ROC-AUC evaluation, the estimation accuracy of the presence or absence of bleeding was 97.7%, which was highly accurate. The classification of bleeding types was also highly accurate, ranging from 96% to 99%.
Below is a graph of the evaluation values.
The image below is an example of inferring the location and type of bleeding. The image on the left is the original CT image. The images in the middle and right show the heat map of the area that the computer is paying attention to as the bleeding point. The heatmap is overlaid on top of the original image.
In this example, there are three correct bleeding points. The AI engine software has detected all of them. The graph shows the accuracy at the time of inference. any indicates the presence or absence of bleeding. The estimated accuracy of bleeding is over 90%. The types of bleeding are more likely to be subdural and intraventricular.
The image below is another example of inferring the location and type of bleeding. The image on the left is the original CT image. The image on the right is a heat map of the area that the computer is paying attention to as a bleeding point. The heatmap is overlaid on top of the original image.
The graph shows the accuracy at the time of inference. any indicates the presence or absence of bleeding. The estimated accuracy of bleeding is over 95%. The type of bleeding has a much higher probability of being intraventricular.
Display of MRI specimen brain by photoreal 3D technology
We performed a three-dimensional display using high-resolution image data (1 pixel has a resolution of 0.1 mm) of the whole brain taken by the 7 Tesla MR device.
Used the following publicly available MRI data.
Edlow, Brian L. et al. (2019), Data from:7 Tesla MRI of the ex vivo human brain at 100 micron resolution, v3, Dryad, Dataset, https://doi.org/10.5061/dryad.119f80q
We use the photorealistic (cinematic) volume rendering method of a new software technology that we have been working on to improve the technology. The resulting image has a representation close to that of the specimen itself.
You can virtually observe cross-sections from all directions without actually slicing the specimen. It has also become possible to depict minute structures such as vessels and nerve fibers.
Below are these amazing images.