A new AI-based pipeline could reduce the time pathologists spend analyzing tissue ingrowth in cerebral aneurysm specimens, according to research presented by Ishaq Ansari of Caldwell University.[1] The automated system achieves high accuracy while reducing the analysis time from days to minutes.

Cerebral aneurysms, which are dilations at weak points in the blood vessels within the brain, pose significant health risks. When these aneurysms rupture, they can lead to hemorrhagic strokes, brain damage, coma, and even death. The Brain Aneurysm Foundation reports that a brain aneurysm ruptures every 18 minutes worldwide, claiming approximately 500 lives annually, with half of the patients being under the age of 50. Among the survivors, 66% experience permanent neurological damage. Current treatment approaches include surgical clipping and endovascular coiling, although the latter often requires retreatment owing to high recurrence rates.

During his presentation, Ansari demonstrated how manual quantification of tissue ingrowth in aneurysmal sacs,  a measure commonly used to assess treatment effectiveness,  can take up to 10 minutes per image. “If we have to quantify 100 images, it will take about three working days continuously,” Ansari said during his presentation. “If you have to quantify tens of thousands of images across multiple labs, it will take almost one year.”

The new AI pipeline, developed in collaboration with Sarder’s computational microscopy imaging lab at the University of Florida, employs a U-Net neural network architecture to segment and quantify tissue ingrowth in histopathological images. The architecture of the model includes 69 layers featuring convolutional layers with batch normalization, ReLU activation functions, and a hierarchical decoding path. Despite working with a relatively small dataset of 64 images, the team employed data augmentation techniques to increase the training set to 200 images.

“One of the key strengths of the U-Net architecture is its ability to perform exceptionally well on small datasets, thanks to its encoder-decoder structure and skip connections that preserve spatial information,” Ansari explained. “Its deep architecture, with batch normalization and multi-scale feature fusion, adapts to heterogeneous tissue patterns.”

Ansari explained that the system achieved promising performance metrics.

“Currently, we have achieved an accuracy of 99%, with precision and recall of 95% and 94%, respectively,” he said. “High precision means fewer false positives, ensuring that only actual ingrown tissue is identified, which is critical for avoiding unnecessary interventions. Meanwhile, a recall of 94% ensures that almost all actual ingrown tissue is detected, minimizing the likelihood of missing important growths that require clinical attention.”

For pathologists concerned about variability in histopathological specimens, the system demonstrated robust performance across different staining conditions and tissue morphologies. The team employed various data augmentation techniques, including flip, rotation, and contrast adjustments, to simulate the range of variations typically encountered in clinical practice.

The pipeline includes sophisticated post-processing steps to ensure clinical usability. “Initially, our focus was on identifying the sac boundary, which was sometimes predicted as incomplete,” Ansari noted in the interview. “To address this, we implemented thresholding and post-processing techniques to reconstruct a complete boundary.”

The team plans to make their tool accessible to the broader pathology community. “The University of Florida has a Computational Microscopy platform where we will first make the model available,” Ansari said. “Yet, deploying it via cloud platforms will help broaden access.”

The team’s next steps include validating the system on larger and more diverse datasets and working closely with pathologists to optimize the interface for clinical workflows. They also plan to explore the applications of this technology to other types of histopathological analyses. “We hope this platform will connect us with more researchers, institutions, and clinical partners, ultimately expanding our access to diverse histopathology datasets,” Ansari concluded.Ishaq Ansari, researcher at Caldwell University, USA.

[1] Ishaq Ansari, AI-based pipeline for automatic quantification of tissue ingrowth in histopathology images of cerebral aneurysm. Presented at SPIE 2025 Digital and Computational Pathology conference, February 18, 2025; San Diego, CA.