VaradPatil8815/wellscanhealthcare-brain-tumor

WellScan Healthcare — Attention ResU-Net Brain Tumor Model

Brain tumor classification + segmentation model trained on MRI scans.
Part of the WellScan Healthcare project — BTech Final Year Project

Model Description

A custom Attention ResU-Net with Classifier — dual-output architecture that simultaneously:

• Classifies the tumor type (4 classes)
• Segments the exact tumor region (pixel-level mask)

Built on U-Net with ResNet skip connections and attention gates integrated into the decoder path.

Model Performance

Metric	Training	Validation
Classification Accuracy	98.73%	97.18%
Segmentation Accuracy	99.48%	99.25%
Mean IoU	0.7312	0.6397
Dice Coefficient	0.8438	0.7769
Total Loss	0.0523	0.0928
Optimal Threshold	0.38	—

Classes

Label	Class
0	Meningioma
1	Glioma
2	Pituitary Tumor
3	No Tumor

Training Details

Parameter	Value
Framework	TensorFlow 2.15 / Keras
Image size	224 × 224 (grayscale)
Training samples	3,727 MRI images
Train/Val/Test split	80/10/10
Optimizer	Adam
LR scheduler	ReduceLROnPlateau
Epochs	50
Batch size	32
Hardware	Kaggle GPU P100 (15GB RAM)

Preprocessing Pipeline

1. Otsu thresholding → binary mask
2. Contour detection → crop to brain region (removes background noise)
3. Resize to 224×224 grayscale
4. Z-score normalization (mean=0, std=1)
5. Data augmentation: rotation, flipping, shifting

How to Use

Load the model

from huggingface_hub import hf_hub_download
from keras.models import load_model
import tensorflow as tf

threshold = 0.38

def mean_iou(y_true, y_pred):
    y_true = tf.cast(y_true, tf.float32)
    y_pred = tf.cast(y_pred >= threshold, tf.float32)
    intersection = tf.reduce_sum(tf.abs(y_true * y_pred))
    union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) - intersection
    return intersection / union

def dice_coefficient(y_true, y_pred, smooth=1e-5):
    y_true = tf.cast(y_true, tf.float32)
    y_pred = tf.cast(y_pred >= threshold, tf.float32)
    intersection = tf.reduce_sum(y_true * y_pred)
    union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred)
    return (2. * intersection + smooth) / (union + smooth)

# Download model from Hugging Face
model_path = hf_hub_download(
    repo_id="varad-patil/wellscanhealthcare-brain-tumor",
    filename="AT_RESu_net_all_STD_0.0927_th=0.38.hdf5"
)

# Load with custom metrics
model = load_model(model_path, custom_objects={
    'mean_iou': mean_iou,
    'dice_coefficient': dice_coefficient
})

Run inference

import cv2
import numpy as np

def preprocess(image_path, size=224):
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    _, binary = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)
    contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    contours = sorted(contours, key=cv2.contourArea, reverse=True)[0]
    x, y, w, h = cv2.boundingRect(contours)
    img = cv2.resize(img[y:y+h, x:x+w], (size, size))
    img = (img - img.mean()) / img.std()
    return np.expand_dims(np.expand_dims(img, axis=-1), axis=0)

class_labels = ['meningioma', 'glioma', 'pituitary tumor', 'noTumor']

img = preprocess("your_mri_scan.jpg")
classification, segmentation_mask = model.predict(img)

predicted_class = class_labels[np.argmax(classification)]
confidence = np.max(classification) * 100
print(f"Prediction: {predicted_class} ({confidence:.2f}%)")

Project Links

• 📂 GitHub: varad-patil/wellscanhealthcare
• 🔗 LinkedIn: Varad Patil