utils.py

# import os
import torch
import resnet
import numpy as np
import tensorflow as tf
# import nibabel as nib
import SimpleITK as sitk
import segmentation_models_3D as sm
from torch import nn
# from ttictoc import tic,toc
from skimage import morphology
from keras import backend as K
from scipy import ndimage as ndi
from keras.models import load_model
from patchify import patchify, unpatchify

# from matplotlib import pyplot as plt
# from matplotlib.widgets import Slider

# Función que retorna modelo 3D U-Net para extracción de cerebro
def import_3d_unet(path_3d_unet):
    # Métricas de desempeño
    def dice_coefficient(y_true, y_pred):
        smoothing_factor = 1
        flat_y_true = K.flatten(y_true)
        flat_y_pred = K.flatten(y_pred)
        return (2. * K.sum(flat_y_true * flat_y_pred) + smoothing_factor) / (K.sum(flat_y_true) + K.sum(flat_y_pred) + smoothing_factor)
    
    # Cargar modelo preentrenado
    # with tf.device('/cpu:0'):
    model = load_model(path_3d_unet, custom_objects={'dice_coefficient':dice_coefficient, 'iou_score':sm.metrics.IOUScore(threshold=0.5)})
    return model
    
    
# Función que caraga imagen en formato nifti, aplica filtro N4 y normaliza imagen
def load_img(path):
    # Lectura de MRI T1 formato nifti
    inputImage = sitk.ReadImage(path, sitk.sitkFloat32)

    return inputImage, sitk.GetArrayFromImage(inputImage).astype(np.float32)

# Función que remueve 
def brain_stripping(inputImage, model_unet):
    """----------------------Preprocesamiento imagen MRI-----------------------"""
    image = inputImage
    
    # N4 Bias Field Correction
    maskImage = sitk.OtsuThreshold(inputImage, 0, 1, 200)
    corrector = sitk.N4BiasFieldCorrectionImageFilter()
    corrected_image = corrector.Execute(image, maskImage)
    log_bias_field = corrector.GetLogBiasFieldAsImage(inputImage)
    corrected_image_full_resolution = inputImage / sitk.Exp(log_bias_field)
    
    #Normalización
    image_normalized = sitk.GetArrayFromImage(corrected_image_full_resolution)
    image_normalized = (image_normalized-np.min(image_normalized))/(np.max(image_normalized)-np.min(image_normalized))
    image_normalized = image_normalized.astype(np.float32)
    
    # Redimención
    mri_image = np.transpose(image_normalized)
    mri_image = np.append(mri_image, np.zeros((192-mri_image.shape[0],256,256,)), axis=0)
    
    # Rotación
    mri_image = mri_image.astype(np.float32)
    mri_image = np.rot90(mri_image, axes=(1,2))

    # Volume sampling
    mri_patches = patchify(mri_image, (64, 64, 64), step=64)

    """--------------------Predicción de máscara de cerebro--------------------"""
    # Máscara de cerebro para cada volúmen
    mask_patches = []

    for i in range(mri_patches.shape[0]):
      for j in range(mri_patches.shape[1]):
        for k in range(mri_patches.shape[2]):
           single_patch = np.expand_dims(mri_patches[i,j,k,:,:,:], axis=0)
           single_patch_prediction = model_unet.predict(single_patch)
           single_patch_prediction_th = (single_patch_prediction[0,:,:,:,0] > 0.5).astype(np.uint8)
           mask_patches.append(single_patch_prediction_th)

    # Conversión a numpy array
    predicted_patches = np.array(mask_patches)

    # Reshape para proceso de reconstrucción
    predicted_patches_reshaped = np.reshape(predicted_patches, 
                                            (mri_patches.shape[0], mri_patches.shape[1], mri_patches.shape[2],
                                             mri_patches.shape[3], mri_patches.shape[4], mri_patches.shape[5]) )

    # Reconstrucción máscara
    reconstructed_mask = unpatchify(predicted_patches_reshaped, mri_image.shape)

    # Suavizado máscara
    corrected_mask = ndi.binary_closing(reconstructed_mask, structure=morphology.ball(2)).astype(np.uint8)

    # Eliminación de volumenes ruido
    no_noise_mask = corrected_mask.copy()
    mask_labeled = morphology.label(corrected_mask, background=0, connectivity=3)
    label_count = np.unique(mask_labeled, return_counts=True)
    brain_label = np.argmax(label_count[1][1:]) + 1

    no_noise_mask[np.where(mask_labeled != brain_label)] = 0

    # Elimicación huecos y hendiduras
    filled_mask = ndi.binary_closing(no_noise_mask, structure=morphology.ball(12)).astype(np.uint8)

    """-------------------------Extracción de cerebro--------------------------"""
    # Aplicar máscara a imagen mri
    mri_brain = np.multiply(mri_image,filled_mask)
    
    return mri_brain

# Función que retorna modelo MedNet
def create_mednet(weight_path, device_ids): 
    # Clase para agregar capa totalmente conectada
    class simci_net(nn.Module):
        def __init__(self):
            super(simci_net, self).__init__()
            
            self.pretrained_model = resnet.resnet50(sample_input_D=192, sample_input_H=256, sample_input_W=256, num_seg_classes=2, no_cuda = False)
            self.pretrained_model.conv_seg = nn.Sequential(nn.AdaptiveMaxPool3d(output_size=(1, 1, 1)),
                                                           nn.Flatten(start_dim=1))
            
            
        def forward(self, x):
            x = self.pretrained_model(x)   
            
            return x
    
    # Path con pesos preentrenados
    weight_path = weight_path
    
    # Lista de GPUs para utilizar
    device_ids = device_ids 
    
    # Generar red
    simci_model = simci_net()
    
    # Distribuir en varias GPUs
    simci_model = torch.nn.DataParallel(simci_model, device_ids = device_ids)
    simci_model.to(f'cuda:{simci_model.device_ids[0]}')
    
    # Diccionario state
    net_dict = simci_model.state_dict()
    
    # Cargar pesos
    weight = torch.load(weight_path, map_location=torch.device(f'cuda:{simci_model.device_ids[0]}'))
    
    # Transferencia de aprendizaje
    pretrain_dict = {}
    
    for k, v in weight['state_dict'].items():
        if k.replace("module.", "module.pretrained_model.") in net_dict.keys():
            pretrain_dict[k.replace("module.", "module.pretrained_model.")] = v     
    
    # pretrain_dict = {k.replace("module.", ""): v for k, v in weight['state_dict'].items() if k.replace("module.", "") in net_dict.keys()}
    net_dict.update(pretrain_dict)
    simci_model.load_state_dict(net_dict)
    
    # Bloqueo de parametros mednet
    for param in simci_model.module.pretrained_model.parameters():
        param.requires_grad = False
        
    simci_model.eval() # Modelo en modo evaluación
        
    return simci_model

# Función que extrae características de cerebro
def get_features(brain, mednet_model):
    with torch.no_grad():
        # Convertir a tensor
        data = torch.from_numpy(np.expand_dims(np.expand_dims(brain,axis=0), axis=0))
        
        # Enviar imagen a GPU
        data = data.to(f'cuda:{mednet_model.device_ids[0]}')
        
        # Extraer Características
        features = mednet_model(data) # Forward
        features = features.cpu().numpy()
    
        torch.cuda.empty_cache()
    return features