Inspiration
Metastasis causes over 90% of cancer-related deaths as of 2025. Despite this high rate, there is no effective system for detecting circulating tumor cells (CTCs), which drive metastasis. Traditional methods depend on expensive, time-consuming staining and manual analysis. I saw this gap and wanted to create a noninvasive, AI-driven framework that could identify and predict metastasis before it happens.
What it does
MetastaScan AI is a complete deep learning system for the early detection of metastasis. It automates the identification of CTCs in liquid biopsy whole-slide imaging (WSI) and predicts both the risk of metastasis and where it is likely to spread using multimodal data fusion.
How we built it
The system has two main parts:
- CTC Detection: We created a new CNN called W17NET_v2, inspired by U-Net. It uses attention gates, skip connections, and atrous convolutions to segment CTCs from WSI. • We achieved 98.7% accuracy on a Caltech dataset with over 120,000 annotated biopsy images. • We used OpenCV, TensorFlow, and NumPy for preprocessing, patching, augmentation, and data formatting.
- Prognosis Models: • We combined extracted morphological, topological, and intensity-based CTC features with patient demographics and tumor data. • A Random Forest classifier predicts metastasis risk with 97.3% accuracy. • An Artificial Neural Network (ANN) predicts the metastasis site with 92.7% accuracy.
The final output includes the segmented image, feature metrics, risk score, and predicted metastasis site.
Challenges we ran into
• We had to deal with data imbalance since CTCs account for less than 1% of total blood cells. • We worked on reducing noise and artifacts in microscopy images. • We optimized performance with limited computing resources while keeping accuracy intact. • We created a pipeline that integrates different data types, including images, genetics, and demographics.
Accomplishments that we're proud of
• We developed a custom CNN architecture that exceeds the performance of standard biomedical models like U-Net. • We achieved nearly human-level accuracy in detecting CTCs and predicting metastasis. • We showed the potential for real-time, noninvasive cancer prognosis with minimal human input.
What we learned
• Bringing together computer vision and structured clinical inference can lead to new levels of diagnostic accuracy. • AI can shift medicine from reactive treatments to proactive prevention. • Strong preprocessing and model explainability are just as important as accuracy in clinical AI design.
What's next for MetastaScan AI
• We plan to expand datasets with various cancer types and imaging sources. • We want to integrate live-cell and unstained microscopy for real-time analysis. • We aim to deploy it as a cloud-based clinical decision support tool integrated with hospitals. • We will conduct multi-center validation studies for regulatory readiness and clinical trials.
Log in or sign up for Devpost to join the conversation.