Model Card for Brain Tumor MRI Segmentation Model

A deep-learning model for brain tumor segmentation in MRI scans using a U-Net-based convolutional neural network. The model outputs a binary mask identifying tumor regions and is intended for research and educational use only.

Model Details

Model Description

This model performs pixel-wise segmentation of brain tumors from MRI slices.
It was developed as part of an academic computer vision project focusing on medical image segmentation.
The architecture is a U-Net implemented in TensorFlow/Keras, trained on the Brain Tumor MRI dataset by Jun Cheng (2017), available on Figshare (DOI: 10.6084/m9.figshare.1512427.v8). The dataset contains MRI scans labeled for tumor segmentation and is commonly used in medical image analysis research.

Developed by: Yara Elshehawi
Model type: U-Net convolutional neural network for image segmentation
License: Apache-2.0

Model Sources

Uses

Direct Use

The model can be used to:

Segment tumor regions in MRI slices
Support research in medical image processing
Serve as a reference implementation for U-Net segmentation tasks
Be integrated into pipelines requiring binary mask prediction

Downstream Use (Optional)

Fine-tuning on other medical segmentation datasets
Adaptation to multi-class tumor segmentation
Integration into educational tools or visualization dashboards

Out-of-Scope Use

The model must not be used for:

Real clinical diagnosis or medical decision-making
Automated medical treatment recommendations
High-stakes deployment on real patient data without proper validation and certification

Bias, Risks, and Limitations

The model is trained on a limited dataset, which may not generalize to all MRI scanners, protocols, or patient populations.
Segmentation quality may vary depending on image contrast, noise, or MRI orientation.
Not suitable for real-world clinical environments without extensive validation.

Recommendations

Users should:

Avoid using the model on out-of-domain MRI scans
Conduct their own evaluation before using it in any research pipeline
Be aware that predictions may include false positives/negatives

How to Get Started with the Model

You can use this model in two ways:

Load it directly from Hugging Face

from tensorflow.keras.models import load_model
from keras import backend as K
import tensorflow as tf
import numpy as np
from tensorflow.keras.preprocessing import image


# Dice Coefficient (required for custom_objects)

def dice_coef(y_true, y_pred):
    y_true = K.cast(y_true, 'float32')
    y_pred = K.cast(y_pred, 'float32')
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersect = K.sum(y_true_f * y_pred_f)
    return (2. * intersect + K.epsilon()) / (
        K.sum(y_true_f) + K.sum(y_pred_f) + K.epsilon()
    )

# Load the model from Hugging Face
model = load_model(
    "yaraa11/brain-tumor-segmentation",
    custom_objects={"dice_coef": dice_coef},
    compile=False
)

# (Optional) Recompile
model.compile(
    optimizer="adam",
    loss="binary_crossentropy",
    metrics=[
        "accuracy",
        tf.keras.metrics.Precision(),
        tf.keras.metrics.Recall(),
        dice_coef
    ]
)

# Load and preprocess input image
img = image.load_img(
    "path_to_image.png",
    target_size=(224, 224),
    color_mode="grayscale"
)
img = image.img_to_array(img) / 255.0
img = np.expand_dims(img, axis=0)

# Predict mask
prediction = model.predict(img)

# Convert to binary mask
binary_mask = (prediction > 0.5).astype(np.float32)

Or load it locally if you downloaded the .keras file

from tensorflow.keras.models import load_model
import numpy as np
from tensorflow.keras.preprocessing import image

model = load_model(
    "./model.keras",
    custom_objects={"dice_coef": dice_coef},
    compile=False
)


# Load and preprocess input image
img = image.load_img(
    "path_to_image.png",
    target_size=(224, 224),
    color_mode="grayscale"
)
img = image.img_to_array(img) / 255.0
img = np.expand_dims(img, axis=0)

# Predict mask
prediction = model.predict(img)

# Convert to binary mask
binary_mask = (prediction > 0.5).astype(np.float32)

Training Details

Training Data

The model was trained on a publicly available brain MRI dataset. All images were preprocessed to ensure consistency across the dataset, including:

Resizing to 256 × 256
Normalizing intensities to [0, 1]
Converting images to single-channel (grayscale) format
Splitting the dataset into training, validation, and testing subsets

Training Procedure

Preprocessing

Before training, each MRI slice was:

Normalized to [0, 1]
Resized to 256 × 256
Converted to grayscale (single-channel) format

Training Hyperparameters

Loss function: Binary Cross-Entropy
Optimizer: Adam
Learning rate: 0.001
Batch size: 16
Epochs: 40

The model was trained end-to-end using TensorFlow/Keras.

Speeds, Sizes, Times

Model size: 6 MB
Training time: Approximately 1–3 hours on a single Colab GPU (e.g., T4)
Number of parameters: 487,009 (~1.86 MB)

Evaluation

Testing Data, Factors & Metrics

Testing Data

Evaluation was conducted on a held-out portion of the MRI dataset that was not used during training.
All preprocessing steps were identical to those used during training.

Factors

Model performance may vary depending on:

Tumor size and location
MRI contrast variations
Image noise, artifacts, or motion
Differences in scanning equipment or acquisition protocols

Metrics

The model was evaluated using standard medical image segmentation metrics:

Accuracy
Precision
Recall
Intersection over Union (IoU)
Dice Coefficient (primary metric for tumor segmentation)

Results

Dice: 0.7019
IoU: 0.4996
Precision: 0.7860
Recall: 0.7714

Summary

The model performs strongly on the provided dataset and reliably segments tumor regions.
However, performance may degrade on MRI scans from different sources or with significantly different characteristics.

Model Examination

The U-Net architecture uses skip connections that allow fine-grained spatial information to be preserved, improving boundary definition and tumor localization.
Performance is influenced heavily by input image quality and consistent preprocessing.

Environmental Impact

Hardware Type: Single NVIDIA T4 GPU (Google Colab)
Training Duration: Approximately 8 hours
Cloud Provider: Google Colab
Estimated Carbon Emissions: <0.2 kg CO₂ (low-impact training)

Technical Specifications

Model Architecture and Objective

Architecture: U-Net encoder–decoder CNN
Objective: Binary semantic segmentation (tumor vs. non-tumor)
Activation: Sigmoid output for pixel-wise classification
Framework: TensorFlow/Keras

Compute Infrastructure

Hardware

Google Colab GPU (T4 / RTX class)

Software

Python 3.10
TensorFlow / Keras
NumPy, OpenCV, SciPy

Citation ()

@article{Cheng2017,
author = "Jun Cheng",
title = "{brain tumor dataset}",
year = "2017",
month = "4",
url = "https://figshare.com/articles/dataset/brain_tumor_dataset/1512427",
doi = "10.6084/m9.figshare.1512427.v8"
}

Glossary

Dice Coefficient:
A segmentation metric measuring overlap between predicted and ground-truth masks. Ranges from 0 (no overlap) to 1 (perfect overlap). Commonly used in medical imaging.

IoU (Intersection over Union):
A metric evaluating how well predicted regions match the true regions by dividing the intersection area by the union area.

Segmentation Mask:
A binary image where each pixel is labeled as either tumor (1) or background (0).

U-Net:
A convolutional neural network architecture with an encoder–decoder structure and skip connections, widely used for biomedical image segmentation.

More Information

For additional details, sample predictions, or troubleshooting, please refer to the repository’s README or open a discussion on the Hugging Face model page.

Model Card Authors

Created by Yara Elshehawi, as part of a medical image segmentation project focused on brain tumor detection using MRI data.

Contact Information

If you have any questions, encounter issues, or are interested in collaboration, feel free to reach out through any of the following channels:

Hugging Face Repository Discussion Board: Visit the "Community" section of the model page to start a discussion or post an issue.
Portfolio Website: You can also contact me directly through my portfolio website for additional inquiries.

Looking forward to hearing from you!

Downloads last month: 76

yaraa11
/

brain-tumor-segmentation