BrainConnect - ASD
Getting an autism diagnosis takes years. Families wait while kids miss the window where early support matters most. BrainConnect cuts that wait by analyzing brain scans and flagging signs of ASD in under a second. The hard part isn't the AI — it's making it work across hospitals. Brain scanners at different sites produce slightly different data, and most models fail the moment they see a new scanner. We fixed this by training four models, each blind to one hospital's data, then letting them vote at prediction time. If they agree on a scan none of them trained on, the pattern is real. This hit 78.7% accuracy across 529 unseen patients. The AMD MI300X ran our language model fine-tune 4.5x faster than an A100 — turning raw predictions into plain-English clinical summaries a doctor can actually hand to a family. The system also shows which brain connections drove each prediction on a 3D brain map, so clinicians aren't just trusting a black box. Supports multiple brain atlas formats so it works with however your hospital preprocessed the data. Upload a scan on Hugging Face Spaces, get a verdict in under a second.
Category tags:
.png&w=96&q=75){kind=small}
"🤖 BrainConnect-ASD — Project Evaluation Project Summary BrainConnect detects Autism Spectrum Disorder (ASD) from brain scans (resting-state fMRI) in under a second. The critical innovation: works across different hospitals/scanners — most models fail when they see a new scanner because they learn "scanner fingerprint" artifacts instead of biological markers. The Solution: Train four models, each blind to one hospital's data. At prediction time, if all four agree on a scan none of them trained on, the pattern is real. This achieved 78.7% accuracy across 529 unseen patients. Key Features: - Adversarial Gradient Reversal Layer (GRL) to ignore scanner artifacts - End-to-end pipeline: fMRI → ASD probability + saliency maps + clinical reports - GCN inference < 20ms - AMD MI300X: Qwen2.5-7B fine-tuning 4.5x faster than A100 - 3D brain map showing which connections drove each prediction Results: - Leave-One-Site-Out (LOSO) validation on ABIDE I: 78.7% accuracy (529 subjects, 20 institutions) - Model: 105K parameters (lightweight) --- 📊 Detailed Scores Application of Technology — 🚀 5/5 This project demonstrates exceptional technical depth: - Adversarial GCN: Uses Gradient Reversal Layer to penalize scanner identification — forces learning only autism-relevant patterns - Cross-site generalization: The core innovation — 4 models, each blind to one site, vote at prediction time - AMD MI300X integration: - Qwen2.5-7B in bf16 (no quantization) - 60 parallel training sessions for LOSO grid search - 4.5x faster than A100 - Lightweight model: 105K parameters, <20ms inference - Full pipeline: BOLD time series → GCN → probability + saliency maps + clinical reports - HuggingFace Spaces demo: Available (though sleeping) The technical solution to "scanner fingerprint" leakage is sophisticated and addresses a real medical AI problem. --- Presentation — 🚀 4/5 Strengths: - Excellent GitHub README with clear problem statement - Detailed architecture with data flow diagram - Table of results by site (NYU, USM, UCLA, UM) - Clear explanation of the adversarial approach - AMD performance metrics (4.5x faster than A100) - Live demo on HuggingFace Spaces Weaknesses: - HuggingFace Space is sleeping (inactive) — couldn't test - Would have liked to see actual 3D brain map visualization - Video demo exists but I can't view --- Business Value — 🚀 5/5 Massive impact: - Autism diagnosis wait times: years → under 1 second - Early intervention window: kids miss this while waiting for diagnosis - Cross-site generalization: works across different hospitals — critical for real deployment The problem is huge: - Families wait years for diagnosis - Early support matters most during developmental windows - Current diagnostic process is slow and backlogged Target market: - Hospitals (diagnostic support) - Research institutions - Clinicians (decision support, not replacement) Value proposition: - Under 1 second diagnosis - Works on any scanner (cross-site) - Explainable: 3D brain map shows which regions drove the prediction - Clinical reports in plain English This could genuinely change lives by speeding up diagnosis. --- Originality — 🚀 5/5 Highly original: - Adversarial deconfounding: Gradient Reversal Layer to ignore scanner artifacts — clever ML technique - Four-model voting: Novel ensemble approach for cross-site generalization - Graph-based brain analysis: GCN on functional connectivity is domain-specific and sophisticated - Zero-shot on unseen sites: The LOSO validation shows it actually generalizes Not just another AI project: - Medical domain (healthcare) - Solves a specific, real problem (scanner variance) - Has clear social impact (faster diagnosis) This stands out from typical hackathon projects because it's a genuine healthcare application with real-world impact. --- 🎯 Final Verdict This is a championship-level project with genuine social impact. The adversarial approach to cross-site generalization is sophisticated ML engineering, and the 78.7% accuracy on unseen patients is an "honest" metric (unlike other papers that use same-site splits claiming 85%+). Strengths: Real healthcare impact, sophisticated technique (GRL), AMD speed advantage, explainable (3D brain maps), cross-site generalization Weaknesses: Demo sleeping, couldn't test directly This is exactly the kind of project that wins hackathons: solves a real problem, has clear impact, technical depth. "
Sanem Avcil
"The window of work was before hackathon started but other than that excellent."
.png&w=96&q=75){kind=small}
Malini Bhandaru
For Innovators & Creators