5️⃣Finally, MELD Graph was a huge team effort with @konradwagstyl.bsky.social, @sophieadlerwagstyl.bsky.social, @hannahspitzer.
and our MELD consortium (@drfelicedarco.bsky.social @nathantcohen.bsky.social @metricsemma.bsky.social @kirstiejane.bsky.social @lzjwilliams.bsky.social + many others!)

4️⃣ MELD Graph is available open-source on our GitHub (github.com/MELDProject) and can be installed on Linux, Mac and Windows. Check out our YouTube tutorials (www.youtube.com/@MELDproject...)!

3️⃣ Notably, MELD Graph detected 64% of lesions previously missed by radiologists. The MELD reports below show two independent test patients with FCD that were missed by 5/5 expert radiologists.

2️⃣ Incorporating whole brain context significantly boosted model specificity – fewer false positives mean the positive predictive is significantly lower than a baseline Multilayer Perceptron (MLP). Also, the predictions generally look much nicer!

And we’ve wrapped it into one neat package.
With one command, MELD Graph processes MRI scans to create an interpretable report that highlights lesion location, describes the lesional features and shares a nicely calibrated confidence score.

1️⃣MELD Graph uses a graph convolutional neural network to segment FCD lesions on the cortical surface. We trained it using the Multicentre Epilepsy Lesion Detection project’s FCD cohort, with 703 epilepsy patients and 482 controls from 23 hospitals around the world.