Beijuka/maternal-ultrasound-funnel-segmentation

Funnel Region Segmentation Modal for Maternal Ultrasound Image Preprocessing

Overview

This project provides a deep learning segmentation model for detecting the funnel-like region of interest (ROI) in maternal ultrasound images. The model is designed as a preprocessing component for downstream maternal healthcare AI pipelines and ultrasound image analysis workflows.

The main objective is to automatically isolate the clinically relevant ultrasound region while suppressing irrelevant image areas and artifacts. The segmented region can then be used to generate ROI-focused ultrasound images where regions outside the detected funnel area are masked to black.

---

Model Architecture

The model was trained using the Ultralytics YOLO segmentation framework.

Base Model

• YOLO11n-seg

Task

• Image Segmentation

Framework

• Ultralytics YOLO

Input Resolution

• 640 × 640

---

Dataset

Dataset Description

The dataset consists of manually annotated maternal ultrasound images with segmentation masks outlining the funnel-shaped ultrasound viewing region.

Total Images

• 161 labeled ultrasound images

Dataset Split

The dataset was randomly shuffled and split into training, validation, and testing sets.

Split	Percentage	Approximate Images
Train	70%	112
Validation	15%	24
Test	15%	25

---

Training Configuration

Training Setup

from ultralytics import YOLO

model = YOLO("yolo11n-seg.pt")

results = model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=8,

    # Conservative medical imaging augmentations
    degrees=10,
    scale=0.1,
    fliplr=0.5,

    mosaic=0.0
)

---

Data Augmentation Strategy

Only conservative augmentations were applied to preserve anatomical and ultrasound image realism.

Augmentation	Value
Rotation	±10°
Scaling	0.1
Horizontal Flip	0.5
Mosaic	Disabled

Mosaic augmentation was disabled to avoid generating unrealistic ultrasound compositions.

---

Model Performance

Evaluation was performed on the held-out test set.

Segmentation Metrics

Metric	Score
Mask mAP@50	0.995
Mask mAP@50-95	0.995
Precision	0.998
Recall	1.000

Detection Metrics

Metric	Score
Box mAP@50	0.995
Box mAP@50-95	0.995

---

Inference Speed

Stage	Time per Image
Preprocess	5.4 ms
Inference	19.0 ms
Postprocess	3.5 ms

---

Sample Validation Output

Class     Images  Instances      Box(P)      R      mAP50  mAP50-95     Mask(P)      R      mAP50  mAP50-95

all         25         25        0.998      1.0      0.995      0.995      0.998      1.0      0.995      0.995

---

Training Metrics

The repository also includes:

• results.csv containing epoch-level training metrics
• results.png showing training and validation performance curves

Example logged metrics include:

Metric	Description
train/box_loss	Bounding box loss
train/seg_loss	Segmentation loss
metrics/mAP50(M)	Segmentation mAP@50
metrics/mAP50-95(M)	Segmentation mAP@50-95
val/seg_loss	Validation segmentation loss

---

Running Inference

Install Dependencies

pip install ultralytics

Load the Model

from ultralytics import YOLO

model = YOLO("best.pt")

Run Prediction

results = model.predict(
    source="image.png",
    conf=0.25
)

---

Repository Contents

File	Description
best.pt	Trained segmentation model
results.csv	Training metrics log
results.png	Training curves
data.yaml	Dataset configuration
README.md	Project documentation

---

Limitations

• The dataset size is relatively small (161 images)
• Performance may vary across ultrasound devices and scanning protocols

Citation

If you use this work, please cite:

@misc{maternal_ultrasound_funnel_segmentation,
  title={Funnel Region Segmentation for Maternal Ultrasound Image Preprocessing},
  author={Beijuka Bruno},
  year={2026}
}