I love this quote from Benjamin Franklin: ‘Either write things worth reading, or do things worth writing.’ @upenn.edu

7/
We tested our method on two datasets:
- HASTE images.
- EPI scans.
And showed that it reaches State-of-the-art performance, especially in younger fetuses. Also, our model is contrast agnostic; it generalizes to various modalities. You can find our preprint at arxiv.org/pdf/2410.20532

6/
Testing: Step 2 (Fine-Level)
- Model B handles mid-sized patches (96³) on the cropped volume. Same for model C with 64³ windows.
- The majority voting across A, B, and C defines consistent regions likely containing the brain.
- Model D refines the final binary mask to avoid edge effects.

4/
To tackle maternal tissues that usually confuse U-Nets, we train 4 U-Nets:
- Each is optimized for different patch sizes.
- Synthetic images include full, partial, and absent brains.
This multi-scale approach prepares us to handle complex scenarios during testing.

3/
Our synthesizer has two components:
- One controls the shape of the brain (applied on labels 1 to 7)
- One manages the background (label 0 and labels 8 to 24).

Separate parameters for each category allow us to have fine control over the variability of the shapes (e.g., warping, scaling, noise)

2/
During training, we augment label maps with random background shapes:
- A big ellipse (womb-like).
- Contours inside/outside the ellipse.
- Synthetic “sticks” and “bones” mimicking maternal anatomy.
This creates diverse and realistic label maps.

🧵 1/ Do you have limited annotations and need a robust fetal brain extraction model with endless training data?
We introduce Breadth-Fine Search (BFS) and Deep Focused Sliding Window (DFS): a framework trained on infinite synthetic images derived from a small set of annotated seeds (label maps).

4/
To tackle maternal tissues that usually confuse U-Nets, we train 4 U-Nets:
- Each is optimized for different patch sizes.
- Synthetic images include full, partial, and absent brains.
This multi-scale approach prepares us to handle complex scenarios during testing.

3/
Our synthesizer has two components:
- One controls the shape of the brain (applied on labels 1 to 7)
- One manages the background (label 0 and labels 8 to 24).

Separate parameters for each category allow us to have fine control over the variability of the shapes (e.g., warping, scaling, noise)

2/
During training, we augment label maps with random background shapes:
- A big ellipse (womb-like).
- Contours inside/outside the ellipse.
- Synthetic “sticks” and “bones” mimicking maternal anatomy.
This creates diverse and realistic label maps.