app.py

import torch
import torch.nn as nn
import torchvision.transforms as transforms
import gradio as gr
import numpy as np
from PIL import Image
import cv2

############################################
# ========== ORIGINAL TRAINING UNET =========
############################################

class DoubleConv(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.conv_op = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
            nn.ReLU(inplace=True)
        )

    def forward(self, x):
        return self.conv_op(x)


class DownSample(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.conv = DoubleConv(in_channels, out_channels)
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)

    def forward(self, x):
        down = self.conv(x)
        p = self.pool(down)
        return down, p


class UpSample(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.up = nn.ConvTranspose2d(in_channels, in_channels//2, kernel_size=2, stride=2)
        self.conv = DoubleConv(in_channels, out_channels)

    def forward(self, x1, x2):
        x1 = self.up(x1)
        x = torch.cat([x1, x2], 1)
        return self.conv(x)


class UNet(nn.Module):
    def __init__(self, in_channels=3, num_classes=1):
        super().__init__()

        self.down_convolution_1 = DownSample(in_channels, 64)
        self.down_convolution_2 = DownSample(64, 128)
        self.down_convolution_3 = DownSample(128, 256)
        self.down_convolution_4 = DownSample(256, 512)

        self.bottle_neck = DoubleConv(512, 1024)

        self.up_convolution_1 = UpSample(1024, 512)
        self.up_convolution_2 = UpSample(512, 256)
        self.up_convolution_3 = UpSample(256, 128)
        self.up_convolution_4 = UpSample(128, 64)

        self.out = nn.Conv2d(in_channels=64, out_channels=num_classes, kernel_size=1)

    def forward(self, x):
        down_1, p1 = self.down_convolution_1(x)
        down_2, p2 = self.down_convolution_2(p1)
        down_3, p3 = self.down_convolution_3(p2)
        down_4, p4 = self.down_convolution_4(p3)

        b = self.bottle_neck(p4)

        up_1 = self.up_convolution_1(b, down_4)
        up_2 = self.up_convolution_2(up_1, down_3)
        up_3 = self.up_convolution_3(up_2, down_2)
        up_4 = self.up_convolution_4(up_3, down_1)

        return self.out(up_4)


############################################
# ========== LOAD MODEL ====================
############################################

device = torch.device("cpu")

model = UNet(in_channels=3, num_classes=1)
model.load_state_dict(torch.load("my_checkpoint.pth", map_location=device))
model.eval()

############################################
# ========== TRANSFORM =====================
############################################

transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor()
])

############################################
# ========== DICE FUNCTION =================
############################################

def dice_coefficient(pred, target, epsilon=1e-7):
    pred = (pred > 0.5).float()
    intersection = (pred * target).sum()
    union = pred.sum() + target.sum()
    return ((2. * intersection + epsilon) / (union + epsilon)).item()

############################################
# ========== TIFF SAFE LOADER ==============
############################################

def load_image(file):
    img = Image.open(file.name)
    img_np = np.array(img)

    # Handle 16-bit TIFF
    if img_np.dtype == np.uint16:
        img_np = (img_np / 256).astype(np.uint8)

    img_pil = Image.fromarray(img_np).convert("RGB")
    return img_pil, img_np

############################################
# ========== PREDICTION ====================
############################################

def predict(image_file, mask_file=None):

    if image_file is None:
        return None, "Please upload an image."

    image_pil, original_np = load_image(image_file)
    input_tensor = transform(image_pil).unsqueeze(0)

    with torch.no_grad():
        output = model(input_tensor)
        output = torch.sigmoid(output)

    pred_mask = output.squeeze().numpy()
    pred_binary = (pred_mask > 0.5).astype(np.uint8)

    # Resize mask back to original size
    pred_resized = cv2.resize(
        pred_binary,
        (original_np.shape[1], original_np.shape[0])
    )

    overlay = original_np.copy()
    overlay[pred_resized == 1] = [255, 0, 0]

    if mask_file is not None:
        mask_pil, _ = load_image(mask_file)
        mask_tensor = transform(mask_pil.convert("L"))
        dice = dice_coefficient(torch.tensor(pred_mask), mask_tensor)
        return overlay, f"Dice Score: {round(dice,4)}"

    return overlay, "Prediction complete."

############################################
# ========== GRADIO UI =====================
############################################

description = """
# 🧠 Brain Tumor Segmentation (UNet)

Dataset used for training:
https://www.kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation

Upload a `.tif` MRI image.
Optionally upload the ground-truth mask to compute Dice score.
"""

demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.File(file_types=[".tif", ".tiff", ".png", ".jpg"], label="Upload MRI Image"),
        gr.File(file_types=[".tif", ".tiff"], label="Optional Ground Truth Mask")
    ],
    outputs=[
        gr.Image(label="Predicted Overlay"),
        gr.Textbox(label="Result")
    ],
    title="UNet Brain Tumor Segmentation",
    description=description
)

if __name__ == "__main__":
    demo.launch()