app.py

import torch
import numpy as np
import pydicom
import gradio as gr
from torchvision import transforms
from PIL import Image

# Load your PyTorch model
model = torch.hub.load('pytorch/vision:v0.6.0', 'resnet18', pretrained=True).eval()


# Define a function to preprocess the DICOM
def preprocess_dicom_to_image(dicom_path):
    # Load DICOM file
    dicom = pydicom.dcmread(dicom_path)
    image = dicom.pixel_array  # Extract image data
    # Normalize to [0, 1] and convert to PIL Image for transforms
    image = (image - np.min(image)) / (np.max(image) - np.min(image))
    return Image.fromarray((image * 255).astype(np.uint8))

# Define a function to preprocess the DICOM
def preprocess_dicom(dicom_path):
    image = preprocess_dicom_to_image(dicom_path)
    # Apply transforms
    transform = transforms.Compose([
        transforms.Resize((224, 224)),  # Resize to model's input size
        transforms.ToTensor(),
    ])
    return transform(image).unsqueeze(0)  # Add batch dimension

# Prediction function
def predict_dicom(dicom_file):
    # Preprocess
    input_tensor = preprocess_dicom(dicom_file.name)
    # Inference
    with torch.no_grad():
        output = model(input_tensor)
    # Convert output tensor to image (dummy example, replace as needed)
    output_image = output.squeeze().numpy()
    output_image = (output_image - np.min(output_image)) / (np.max(output_image) - np.min(output_image)) * 255
    output_image = Image.fromarray(output_image.astype(np.uint8))
    return output_image

# Prediction function
def predict_dicom2(dicom_file):
    return preprocess_dicom_to_image(dicom_file)

# Create Gradio interface
interface = gr.Interface(
    fn=predict_dicom2,
    inputs=gr.File(label="Upload DICOM File"),
    outputs="image",
    title="DICOM Image Prediction"
)

# Launch the Gradio app
interface.launch()