OncoTriage AMD-Boosted Uncertainty-Aware CT Triage
OncoTriage is a clinical decision support system designed to detect and triage lung nodules in chest X-rays with high reliability. Developed solo for the 2026 AMD Hackathon, it addresses the lack of transparency in automated diagnostics by implementing Bayesian Deep Learning. The system utilizes a Bayesian EfficientNet-B4 backbone. By employing MC Dropout, the model generates a predictive distribution rather than a single point estimate, allowing for the calculation of epistemic uncertainty. This effectively quantifies the model's confidence for every detection. In clinical settings, this allows the system to flag low-confidence predictions for priority human review, reducing the risk of false negatives inherent in standard "black-box" AI. In addition to this, in order to handle the intensive computational requirements of Bayesian inference, OncoTriage is optimized for AMD Instinct MI300X instances. Leveraging AMD’s high-bandwidth memory (HBM3) and the ROCm stack, the system achieves the rapid inference times necessary for real-time clinical triage. The environment is fully containerized via Docker, ensuring seamless scalability across high-performance compute clusters. The Mission: OncoTriage represents a shift toward accountable, transparent AI. By bridging the gap between raw computational power and clinical safety, it provides radiologists with a reliable partner in oncological screening—transforming raw data into uncertainty-aware medical intelligence.
Category tags:
"This is an impressive solo project addressing a critical healthcare need. The use of Bayesian Deep Learning for uncertainty quantification is particularly valuable in clinical settings - it addresses the "black box" problem in medical AI by providing confidence scores for every prediction. The fact that a solo developer built this makes it even more impressive. Application of Technology: 🚀🚀🚀🚀🚀 5 - Bayesian EfficientNet-B4 backbone, MC Dropout for epistemic uncertainty quantification, AMD MI300X optimization with HBM3 and ROCm, Docker containerization. Sophisticated technical implementation. Presentation: 🚀🚀🚀🚀 4 - Clear explanation of the problem and solution. Good demo link. Could have more details on the architecture but covers essentials well. Business Value: 🚀🚀🚀🚀🚀 5 - Lung cancer screening is a massive healthcare need globally. Early detection saves lives. The uncertainty quantification allows doctors to know when to trust the AI and when to review manually - critical for clinical adoption. Originality: 🚀🚀🚀🚀🚀 5 - Bayesian Deep Learning for medical imaging is relatively novel. The focus on uncertainty awareness rather than just accuracy is a important distinction for clinical safety."
Sanem Avcil
For Innovators & Creators