app.py

import gradio as gr
from PIL import Image
from joblib import load
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import efficientnet_b0
import torchvision.transforms as transforms


class MultiModalClassifier(nn.Module):
    def __init__(self, num_classes, num_features):
        super(MultiModalClassifier, self).__init__()
        # Load pre-trained EfficientNet model
        efficientnet = efficientnet_b0(pretrained=True)
        
        # Remove the last classifier layer
        self.efficientnet_features = nn.Sequential(*list(efficientnet.children())[:-1])
        
        # Define additional feature dimensions
        self.age_dim = 1  # assuming age is a single scalar value
        self.anatom_site_dim = 1  # assuming anatomical site is a single scalar value
        self.sex_dim = 1  # assuming sex is a single scalar value
        
        # Fully connected layers for classification
        self.fc1 = nn.Linear(num_features + self.age_dim + self.anatom_site_dim + self.sex_dim, 256)
        self.fc2 = nn.Linear(256, num_classes)
        
        # Dropout layer.
        self.dropout = nn.Dropout(p=0.5)
        
    def forward(self, image, age, anatom_site, sex):
        # Forward pass through the pre-trained EfficientNet model
        image_features = self.efficientnet_features(image)
        image_features = F.avg_pool2d(image_features, image_features.size()[2:]).view(image.size(0), -1)  # Flatten
        
        # Reshape age, anatom_site, and sex tensors
        age = age.view(-1, 1)  # Reshape to [batch_size, 1]
        anatom_site = anatom_site.view(-1, 1)  # Reshape to [batch_size, 1]
        sex = sex.view(-1, 1)  # Reshape to [batch_size, 1]
        # Concatenate image features with additional features
        additional_features = torch.cat((age, anatom_site, sex), dim=1)
        combined_features = torch.cat((image_features, additional_features), dim=1)
        
        # Fully connected layers for classification
        combined_features = F.relu(self.fc1(combined_features))
        combined_features = self.dropout(combined_features)
        output = self.fc2(combined_features) 
        
        return output

# Initialize the model
num_classes = 1  # Assuming binary classification
num_features = 1280  # Number of features extracted by EfficientNet-B0
model = MultiModalClassifier(num_classes, num_features)

# Load the saved model state dictionary
model.load_state_dict(torch.load(r'best_epoch_weights.pth',map_location=torch.device('cpu')))

# Set the model to evaluation mode
model.eval()

# Load the age scaler
age_scaler = load(r'age_approx_scaler.joblib')

# Define transforms for the data (adjust as necessary.)
test_transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
])

diagnosis_map = {0: 'benign', 1: 'malignant'} 

# Define mapping dictionaries for sexes and anatom_sites.
sexes_mapping = {'male': 0, 'female': 1}

# Define mapping dictionary for anatom_site_general.
anatom_site_mapping = {
    'torso': 0,
    'lower extremity': 1,
    'head/neck': 2,
    'upper extremity': 3,
    'palms/soles': 4,
    'oral/genital': 5,
}

def predict(image, age, gender, anatom_site):

    image = Image.fromarray(image)
    # Apply transformations to the image
    image = test_transform(image)
    image = image.float()
    image = image.unsqueeze(0)  # Add batch dimension

    sex = torch.tensor([[sexes_mapping[gender.lower()]]],  dtype=torch.float32)
    anatom_site = torch.tensor([[anatom_site_mapping[anatom_site]]], dtype=torch.float32)

    # Scale the age using the loaded scaler
    scaled_age = age_scaler.transform([[age]])
    # Convert scaled age to a tensor
    age_tensor = torch.tensor(np.array(scaled_age), dtype=torch.float32)

    # Forward pass
    output = model(image, age_tensor, anatom_site, sex)

    # Apply sigmoid to the output (since it's a binary classification)
    output_sigmoid = torch.sigmoid(output)
    # Get the predicted class (0 or 1)
    predicted_class = (output_sigmoid > 0.5).float()

    
    return f"The predicted_class is a {diagnosis_map[int(predicted_class)]}."


description_html = """
Fill in the required parameters and click 'classify'.
"""

example_data = [
    ["ISIC_0000060_downsampled.jpg", 35, "Female", "torso"],
    ["ISIC_0068279.jpg", 45.0, "Female", "head/neck"]
]

inputs = [
    "image",
    gr.Number(label="Age", minimum=0, maximum=120),
    gr.Dropdown(['Male', 'Female'], label="Gender"),
    gr.Dropdown(['torso', 'lower extremity', 'head/neck', 'upper extremity', 'palms/soles', 'oral/genital'], label="Anatomical Site")
]

gr.Interface(
    predict,
    inputs,
    outputs = gr.Textbox(label="Output", type="text"),
    title="Skin Cancer Diagnosis",
    description=description_html,
    allow_flagging='never',
    examples=example_data
).launch()