Train_SwinUnetr.py

#!/usr/bin/env python
# coding: utf-8




import os
import torch
import numpy as np
import pandas as pd
from tqdm import tqdm
import nibabel as nib

import monai
from monai.networks.nets import UNETR, UNet
from monai.utils import first
from monai.data import CacheDataset, DataLoader, Dataset, decollate_batch
from monai.inferers import SliceInferer

from time import sleep

from samunetr.SAMUNETR import SAMUNETR




#Organize Picai Data

data_picai=pd.read_csv('/home/jaalzate/Tartaglia/Prostate_Tartaglia/codes/partition_1.csv')

data_picai["depth"] = data_picai['filepath_t2w_cropped'].apply(lambda path_file: nib.load(path_file).shape[0])
data_picai["heigth"] = data_picai['filepath_t2w_cropped'].apply(lambda path_file: nib.load(path_file).shape[1])
data_picai["weigth"] = data_picai['filepath_t2w_cropped'].apply(lambda path_file: nib.load(path_file).shape[2])
data_picai=data_picai[(data_picai['heigth']!=0) & (data_picai['depth']!=0)]
data_picai=data_picai[(data_picai['heigth']>96) & (data_picai['depth']>96)]



data_picai=data_picai[data_picai['filepath_t2w_cropped'].notna()].reset_index()
data_picai_human=data_picai[data_picai['human_labeled']==1]
data_picai.drop(data_picai_human.index, inplace = True)

data_picai=data_picai[['filepath_t2w_cropped','filepath_adc_cropped','filepath_hbv_cropped','filepath_labelAI_cropped','filepath_seg_zones_cropped','partition']]
data_picai_human=data_picai_human[['filepath_t2w_cropped','filepath_adc_cropped','filepath_hbv_cropped','filepath_label_cropped','filepath_seg_zones_cropped','partition']]



train_picai=data_picai[data_picai['partition']=='tr']
test_picai=data_picai[data_picai['partition']=='dev']

train_picai_human=data_picai_human[data_picai_human['partition']=='tr']
test_picai_human=data_picai_human[data_picai_human['partition']=='dev']




#Organize Prostate158 Data
path='/nvmescratch/ceib/Prostate/input/prostate158/prostate158_train'
train_df_P158=pd.read_csv(os.path.join(path,'train.csv'))
test_df_P158=pd.read_csv(os.path.join(path,'valid.csv'))

columns = ['t2', 'adc', 'dwi','t2_anatomy_reader1', 'adc_tumor_reader1']

for df in [train_df_P158, test_df_P158]:
    for column in columns:
        df[column] = df[column].apply(lambda x: os.path.join(path,x))




#Combine Datasets

train_df = pd.DataFrame({
    't2w': list(train_picai['filepath_t2w_cropped'].values)+list(train_picai_human['filepath_t2w_cropped'].values)+list(train_df_P158['t2'].values),
    'adc': list(train_picai['filepath_adc_cropped'].values) + list(train_picai_human['filepath_adc_cropped'].values)+list(train_df_P158['adc'].values),
    'dwi': list(train_picai['filepath_hbv_cropped'].values) + list(train_picai_human['filepath_hbv_cropped'].values)+list(train_df_P158['dwi'].values),
    'zones': list(train_picai['filepath_seg_zones_cropped'].values) + list(train_picai_human['filepath_seg_zones_cropped'].values)+list(train_df_P158['t2_anatomy_reader1'].values),
    'label': list(train_picai['filepath_labelAI_cropped'].values) + list(train_picai_human['filepath_label_cropped'].values)+list(train_df_P158['adc_tumor_reader1'].values)
})

test_df = pd.DataFrame({
    't2w': list(test_picai['filepath_t2w_cropped'].values) + list(test_picai_human['filepath_t2w_cropped'].values)+list(test_df_P158['t2'].values),
    'adc': list(test_picai['filepath_adc_cropped'].values) +  list(test_picai_human['filepath_adc_cropped'].values)+list(test_df_P158['adc'].values) ,
    'dwi': list(test_picai['filepath_hbv_cropped'].values) + list(test_picai_human['filepath_hbv_cropped'].values)+list(test_df_P158['dwi'].values),
    'zones': list(test_picai['filepath_seg_zones_cropped'].values) + list(test_picai_human['filepath_seg_zones_cropped'].values)+list(test_df_P158['t2_anatomy_reader1'].values),
    'label': list(test_picai['filepath_labelAI_cropped'].values) + list(test_picai_human['filepath_label_cropped'].values)+list(test_df_P158['adc_tumor_reader1'].values)
})




print(train_df.shape)
print(test_df.shape)



def Create_dataloaders(train_df,test_df,cache=False):

    """
    This function is for preprocessing, it contains only the basic transforms, but you can add more operations that you 
    find in the Monai documentation.
    https://monai.io/docs.html
    """

    #set_determinism(seed=0)

    img_columns=["t2","adc","dwi"]#,"adc","dwi"]
    label_column=["label"]
    
    mode=["bilinear","nearest"]#,"bilinear","bilinear","nearest"]#["bilinear","bilinear","bilinear", "nearest"]

    train_files = [{"t2": t2,'adc': adc,'dwi': dwi,"zones":zones, "label": label} for 
                   t2,adc,dwi,zones, label in zip(train_df['t2w'].values,
                                     train_df['adc'].values,
                                     train_df['dwi'].values,
                                     train_df['zones'].values,
                                     train_df['label'].values)]
    test_files = [{"t2": t2,'adc': adc,'dwi': dwi,"zones":zones, "label": label} for 
                   t2,adc,dwi,zones,label in zip(test_df['t2w'].values,
                                     test_df['adc'].values,
                                     test_df['dwi'].values,
                                     test_df['zones'].values,
                                     test_df['label'].values)]
    prob=0.175
    train_transforms = monai.transforms.Compose(
        [
            monai.transforms.LoadImaged(keys=img_columns+label_column+["zones"],reader="NibabelReader",image_only=True),
            monai.transforms.AsDiscreted(keys=label_column,threshold=1), #Convert values greater than 1 to 1
            monai.transforms.EnsureChannelFirstd(keys=img_columns+label_column+["zones"]),
            monai.transforms.AsDiscreted(keys="zones",argmax=False,to_onehot=3),
            monai.transforms.LabelToMaskd(keys="zones",select_labels=[1,2]),
            monai.transforms.ResampleToMatchd(keys=["adc","dwi","zones","label"],key_dst="t2",mode=("bilinear","bilinear","nearest","nearest")),#Resample images to t2 dimension
            monai.transforms.Resized(keys=img_columns+label_column+["zones"],spatial_size=(128,128,-1),mode=("trilinear","trilinear","trilinear","nearest","nearest")),#SAMUNETR: Reshape to have the same dimension
            #monai.transforms.ScaleIntensityd(keys=img_columns,minv=0.0, maxv=1.0),
            monai.transforms.NormalizeIntensityd(keys=img_columns),
            monai.transforms.ConcatItemsd(keys=img_columns+["zones"], name='image', dim=0),
            monai.transforms.ConcatItemsd(keys=label_column, name='label', dim=0),
            #monai.transforms.RandSpatialCropSamplesd(keys=['image','label'],roi_size=[96,96,-1],num_samples=8,random_size=False),#For the other models
            monai.transforms.RandRotate90d(keys=['image','label'],spatial_axes=[0,1],prob=prob),
            monai.transforms.RandZoomd(keys=['image','label'],min_zoom=0.9,max_zoom=1.1,mode=['area' if x == 'bilinear' else x for x in mode],prob=prob),
            monai.transforms.RandGaussianNoised(keys=["image"],mean=0.1,std=0.25,prob=prob),
            monai.transforms.RandShiftIntensityd(keys=["image"],offsets=0.2,prob=prob),
            monai.transforms.RandGaussianSharpend(keys=['image'],sigma1_x=[0.5, 1.0],sigma1_y=[0.5, 1.0],sigma1_z=[0.5, 1.0],sigma2_x=[0.5, 1.0],sigma2_y=[0.5, 1.0],sigma2_z=[0.5, 1.0],alpha=[10.0,30.0],prob=prob),
            monai.transforms.RandAdjustContrastd(keys=['image'],gamma=2.0,prob=prob),
            
        ]
    )
    test_transforms = monai.transforms.Compose(
        [
            monai.transforms.LoadImaged(keys=img_columns+label_column+["zones"],image_only=True),
            monai.transforms.AsDiscreted(keys=label_column,threshold=1), #Convert values greater than 1 to 1
            monai.transforms.EnsureChannelFirstd(keys=img_columns+label_column+["zones"]),
            monai.transforms.AsDiscreted(keys="zones",argmax=True,to_onehot=3),
            monai.transforms.LabelToMaskd(keys="zones",select_labels=[1,2]),
            monai.transforms.ResampleToMatchd(keys=["adc","dwi","zones","label"],key_dst="t2",mode=("bilinear","bilinear","nearest","nearest")),#Resample images to t2 dimensions
            monai.transforms.Resized(keys=img_columns+label_column+["zones"],spatial_size=(128,128,-1),mode=("trilinear","trilinear","trilinear","nearest","nearest")),#SAMUNETR: Reshape to have the same dimension
            #monai.transforms.ScaleIntensityd(keys=img_columns,minv=0.0, maxv=1.0),
            monai.transforms.NormalizeIntensityd(keys=img_columns),
            monai.transforms.ConcatItemsd(keys=img_columns+["zones"], name='image', dim=0),
            monai.transforms.ConcatItemsd(keys=label_column, name='label', dim=0),
            
        ]
    )
    
    
    if cache:
        train_ds = CacheDataset(data=train_files, transform=train_transforms,cache_rate=1.0,num_workers=8,copy_cache=False)#PerSlice(keys='image',transforms=train_transforms),cache_rate=1.0,num_workers=8,copy_cache=False)
        train_loader = DataLoader(train_ds, batch_size=1,shuffle=True)

        test_ds = CacheDataset(data=test_files, transform=test_transforms, cache_rate=1.0,num_workers=8,copy_cache=False)
        test_loader = DataLoader(test_ds, batch_size=1, shuffle=False)

        return train_loader,train_ds, test_loader,test_ds

    else:
        train_ds = Dataset(data=train_files, transform=train_transforms)
        train_loader = DataLoader(train_ds, batch_size=1,shuffle=True)
        test_ds = Dataset(data=test_files, transform=test_transforms)
        test_loader = DataLoader(test_ds, batch_size=1,shuffle=False)

        return train_loader,train_ds, test_loader,test_ds



def train2D(model, data_in, loss, optim, max_epochs, model_dir,device,name, test_interval=1):
    best_metric = -1
    best_metric_epoch = -1
    save_loss_train = []
    save_loss_test = []
    save_metric_train = []
    save_metric_test = []
    train_loader, test_loader = data_in

    for epoch in range(max_epochs):
        print("-" * 10)
        print(f"epoch {epoch + 1}/{max_epochs}")
        model.train()
        train_epoch_loss = 0
        train_step = 0
        epoch_metric_train = 0
        ap_metric_train=0
        with tqdm(train_loader, unit="batch") as tepoch:
            for batch_data in tepoch:
                tepoch.set_description(f"Epoch {epoch+1}")
                train_step += 1
                #To convert images to 2D
                volume_list = list(batch_data["image"])
                label_list = list(batch_data["label"])
                volume=torch.cat(volume_list,axis=-1)
                label=torch.cat(label_list,axis=-1)

                volume = monai.transforms.Transpose((3,0,1,2))(volume)
                label=monai.transforms.Transpose((3,0,1,2))(label)

                #######################################
                volume, labels = (volume.to(device), label.to(device))
                optim.zero_grad()
                outputs = model(volume)#[0]
                
                train_loss = loss(outputs, labels)

                train_loss.backward()
                optim.step()

                train_epoch_loss += train_loss.item()
            
                labels_list = decollate_batch(labels)
                labels_convert = [post_label(label_tensor) for label_tensor in labels_list]
                
                output_list = decollate_batch(outputs)
                output_convert = [post_pred(output_tensor) for output_tensor in output_list]
                
                dice_metric(y_pred=output_convert, y=labels_convert)
                iou_metric(y_pred=output_convert, y=labels_convert)
            
                tepoch.set_postfix(loss=train_loss.item(), dice_score=dice_metric.aggregate(reduction="mean").item())
                sleep(0.1)

            print('-'*20)

            train_epoch_loss /= train_step
            print(f'Epoch_loss: {train_epoch_loss:.4f}')
            save_loss_train.append(train_epoch_loss)
            np.save(os.path.join(model_dir, name+'_loss_train.npy'), save_loss_train)

            epoch_metric_train = dice_metric.aggregate(reduction="mean").item()
            dice_metric.reset()

            print(f'Epoch_metric: {epoch_metric_train:.4f}')
            
            iou_metric_train = iou_metric.aggregate(reduction="mean").item()
            iou_metric.reset()

            print(f'IoU_metric: {iou_metric_train:.4f}')
            

            save_metric_train.append(epoch_metric_train)
            np.save(os.path.join(model_dir, name+'_metric_train.npy'), save_metric_train)

            if (epoch + 1) % test_interval == 0:

                model.eval()
                with torch.no_grad():
                    test_epoch_loss = 0
                    test_metric = 0
                    epoch_metric_test = 0
                    test_step = 0
                    ap_metric=0
                    for test_data in test_loader:
                        test_step += 1
                        test_volume, test_label = (test_data["image"].to(device),test_data["label"].to(device))
                        inferer=SliceInferer(roi_size=(None, None),sw_batch_size=16,spatial_dim=2,cval=-1,progress=False)
                        test_outputs = inferer(test_volume, model)
                        
                        test_loss = loss(test_outputs, test_label)
                        test_epoch_loss += test_loss.item()
                        
                        labels_list = decollate_batch(test_label)
                        labels_convert = [post_label(label_tensor) for label_tensor in labels_list]

                        output_list = decollate_batch(test_outputs)
                        output_convert = [post_pred(output_tensor) for output_tensor in output_list]

                        dice_metric(y_pred=output_convert, y=labels_convert)
                        iou_metric(y_pred=output_convert, y=labels_convert)

                    test_epoch_loss /= test_step
                    print(f'test_loss_epoch: {test_epoch_loss:.4f}')
                    save_loss_test.append(test_epoch_loss)
                    np.save(os.path.join(model_dir, name+'_loss_test.npy'), save_loss_test)

                    epoch_metric_test=dice_metric.aggregate(reduction="mean").item()
                    
                    print(f'test_dice_epoch: {epoch_metric_test:.4f}')
                    print('test_dice_epoch_per_class:',dice_metric.aggregate())
                    
                    iou_metric_test=iou_metric.aggregate(reduction="mean").item()
                    
                    print(f'test_iou_epoch: {iou_metric_test:.4f}')
                    print('test_iou_epoch_per_class:',iou_metric.aggregate())
                    iou_metric.reset()
                    
                    save_metric_test.append(epoch_metric_test)
                    np.save(os.path.join(model_dir, name+'_metric_test.npy'), save_metric_test)
                    dice_metric.reset()
                    if epoch_metric_test > best_metric:
                        best_metric = epoch_metric_test
                        best_metric_epoch = epoch + 1
                        torch.save(model.state_dict(), os.path.join(
                            model_dir, name+"_best_metric_model.pth"))

                    print(
                        f"current epoch: {epoch + 1} current mean dice: {epoch_metric_test:.4f}"
                        f"\nbest mean dice: {best_metric:.4f} "
                        f"at epoch: {best_metric_epoch}"
                    )


    print(
        f"train completed, best_metric: {best_metric:.4f} "
        f"at epoch: {best_metric_epoch}")




#Creating dataloaders

train_loader,train_ds,val_loader,val_ds=Create_dataloaders(train_df,test_df,cache=True)




pin_memory = torch.cuda.is_available()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print(f'Working on device: {device}')


# # Models

# ## Unet with Residual Units

model=monai.networks.nets.SwinUNETR(
    img_size=(128, 128),
    in_channels=5,
    out_channels=2,
    feature_size=48,
    use_checkpoint=False,
    norm_name="instance",
    depths=(2, 2, 2,2),
    spatial_dims=2,
    drop_rate=0.15
).to(device)

# # Training Model


#loss_function = DiceCELoss(to_onehot_y=True, sigmoid=True, squared_pred=True, ce_weight=calculate_weights(1792651250,2510860).to(device))
#loss_function = monai.losses.DiceCELoss(to_onehot_y=True, sigmoid=False,softmax=True,include_background=True)
loss_function = monai.losses.DiceFocalLoss(to_onehot_y=True, sigmoid=False,softmax=True,include_background=True)

torch.backends.cudnn.benchmark = True
optimizer = monai.optimizers.Novograd(model.parameters(), lr=0.001, weight_decay=0.01)





data_in=(train_loader,val_loader)
model_dir='/home/jaalzate/Tartaglia/Prostate_Tartaglia/Paper_Resultados/Results/SwinUnetr/128x128'



post_pred = monai.transforms.Compose(
        monai.transforms.AsDiscrete(argmax=True, to_onehot=2, num_classes=2),
        monai.transforms.KeepLargestConnectedComponent(applied_labels=list(range(1, 2))),
    )

post_label = monai.transforms.AsDiscrete(to_onehot=2)
dice_metric = monai.metrics.DiceMetric(include_background=False, reduction="mean_batch", get_not_nans=False,ignore_empty=True)
iou_metric=monai.metrics.MeanIoU(include_background=False,reduction="mean_batch",get_not_nans=False,ignore_empty=True)

train2D(model, data_in, loss_function, optimizer, 200, model_dir,device=device,name='SwinUnetr_128x128')