Business Context

MedVision AI is building an assistive tool for radiologists to outline tumors in brain MRI scans. Manual segmentation is slow and variable across annotators, so the team wants a model that produces accurate pixel-level masks for review.

Dataset

You are given a medical image segmentation dataset based on multi-modal brain MRI studies. Each study contains four aligned MRI sequences per patient and a pixel-wise tumor mask. The task is to predict a binary segmentation mask for tumor vs background.

• Size: ~12,000 2D slices from 350 patient studies, 4 input channels
• Target: Binary pixel-wise segmentation mask
• Class balance: Highly imbalanced at pixel level; background dominates and many slices contain small tumor regions
• Missing data: No missing labels, but image intensity distributions vary across scanners and hospitals

Success Criteria

A good solution should achieve strong overlap between predicted and true masks, with Dice score >= 0.82 on the held-out test set and stable performance across patients. The model should also avoid missing small tumor regions.

Constraints

• Inference should complete in <150 ms per 2D slice on a GPU for batch review workflows
• The approach should be robust to limited labeled data
• The model should preserve fine spatial detail needed for clinical review

Deliverables

1. Explain what U-Net is and why it is well-suited for medical image segmentation
2. Build a training pipeline for binary tumor segmentation
3. Evaluate the model using segmentation-specific metrics
4. Describe preprocessing and augmentation choices for MRI data
5. Discuss deployment tradeoffs such as latency, memory use, and false negatives