Segment Anything Model (SAM)-powered spotting of lacunes in brain

Kavi Bharathi

Researchers have recently turned to AI-based tools to screen for lacunes, which are small, fluid-filled cavities in the brain serving as key imaging biomarkers for cerebral small vessel diseases (CSVDs). CSVDs are linked to increased risk of stroke, dementia, cognitive decline, and gait impairments. It is therefore important to accurately identify these lacunes for precise diagnoses of these diseases and assess their prognosis.

Although advancements in MRI have improved the detection of lacunes, challenges persist due to their diverse appearance and size, as well as the presence of mimics such as enlarged perivascular spaces and other fluid-filled cavities. These challenges often lead to high false positive rates, and some portions of normal anatomical structures, such as sulci and cerebral ventricles, can resemble lacunes and be misclassified.

To solve this problem, researchers at IISc have turned to a deep learning-based approach. The team, led by Vaanathi Sundaresan at the Department of Computational and Data Sciences (CDS), developed an automated multistage pipeline that functions like a radiologist in examining MRI scans. They trained a prompt generator to detect all potential lacune candidates from 2D slices, and guided by these prompts, employed Meta AI’s Segment Anything Model (SAM) to accurately identify lacunes and distinguish them from mimics. This approach mirrors clinicians’ practice of examining lacunes in 3D scans while ensuring regional and morphologically consistent characteristics across all axes but at a much faster rate.

The model was found to be robust in accurately identifying lacunes in MRI scans under varied imaging conditions. It outperformed existing state-of-the-art methods and showed strong performance even in low-data settings by employing an innovative process called self-distillation, where the model learns from its own predictions.

When it comes to treating conditions like stroke, timing is a crucial factor. Detecting lacunes on time, therefore, can help boost the effectiveness of treatment. In the future, the team aims to investigate population-level associations between CSVD characteristics and clinical outcomes in cerebrovascular and neurodegenerative diseases.

REFERENCE:
Deepika P, Shanker G, Narayanan R, Sundaresan V, Automated detection of lacunes in brain MR images using SAM with robust prompts using self-distillation and anatomy-informed priors, Computers in Biology and Medicine (2025).
https://doi.org/10.1016/j.compbiomed.2025.110806

LAB WEBSITE:
https://sites.google.com/view/biomedia-lab/home