Researchers develop geometric deep learning model with potential to more accurately detect stroke lesions


Researchers in Germany have developed a new deep learning model for detecting stroke lesions, claiming that—by leveraging rich geometric information to segment brain stroke lesion images—the proposed approach may be able to outperform other, more widely used neural network architectures when it comes to accuracy.

Computed tomography-perfusion (CT-P) is one of the most useful imaging modalities during the early stages of an acute stroke, helping physicians to pinpoint the location of a clot and analyse the extent of damage to brain tissue. However, it is challenging to accurately identify segmentation—the outline of stroke lesions—in a CT-P scan, and the final diagnosis depends greatly on the surgeon’s expertise and ability.

To address this issue, scientists have come up with various machine learning models that perform automatic segmentation of CT-P scans. Unfortunately, none of them have reached a level of performance suitable for clinical applications, as per a news release from the International Society for Optics and Photonics (SPIE).

Against this backdrop, a team of researchers from Germany recently developed a new segmentation algorithm for stroke lesions. As reported in their study, which is published in the Journal of Medical Imaging, the team built a geometric deep learning model called ‘Graph Fully-Convolutional Network’ (GFCN). The internal operations performed by their geometric algorithm differ fundamentally from those of more widely used Euclidean models, the release notes. In their study, the researchers explored the benefits and limitations of this alternative approach.

A key advantage of the proposed model is that it can better learn and preserve important features inherent to brain topology. By using a graph-based neural network, the algorithm can detect complex inter-pixel relationships from different angles. This, in turn, enables it to detect stroke lesions more accurately.

In addition, the team adopted ‘pooling’ and ‘unpooling’ blocks in their network structure. The pooling operations, or ‘downsampling’, reduce the overall size of the feature maps extracted by the network from input images. This reduces the computational complexity of the algorithm, enabling the model to extract the most salient features of the CT-P scans. In contrast, the unpooling operations, or ‘upsampling’, revert the pooling operations to help properly localise the detected features in the original image based on contextual cues. By combining these two operations, the network structure can extract richer geometric information.

The team conducted a series of analyses to determine the effect of each component of GFCN on its segmentation performance. They then compared the performance of the proposed algorithm against the state-of-the-art models, all trained using the same public dataset. Interestingly, the release states, although their model used basic unpooling techniques and a simple input configuration, it performed better than the conventional models under most conditions.

Notably, GFCN-8s, with three pooling layers and eight-fold upsampling, achieved a Dice coefficient score—a metric indicating the overlap between the predicted and actual lesion areas—of 0.4553, which is significantly higher than other models. Moreover, the proposed model could adapt to irregular segmentation boundaries better than established, state-of-the-art models.

Overall, the researchers believe the findings of this study showcase the potential of geometric deep learning for segmentation problems in medical imaging. In addition, further research on similar strategies could pave the way for highly accurate models for automatic stroke diagnosis that could improve patient outcomes and save lives, the release concludes.


Please enter your comment!
Please enter your name here