Could not achieve the accuracy on ETIS-LaribPolypDB as mentioned in the corresponding paper

[EthicalAccount]

I was not able to reproduce the results on this particular dataset that were provided in the research paper.

Everything was kept on default settings with parameters double checked but still could not achieve the desired results

Dice Score
Proposed: 0.9324
Achieved: 0.8485

# Importing the necessary libraries

  import tensorflow as tf
  import albumentations as albu
  import numpy as np
  import gc
  import pickle
  import matplotlib.pyplot as plt
  from keras.callbacks import CSVLogger
  from datetime import datetime
  from sklearn.model_selection import train_test_split
  from sklearn.metrics import jaccard_score, precision_score, recall_score, accuracy_score, f1_score
  from ModelArchitecture.DiceLoss import dice_metric_loss
  from ModelArchitecture import DUCK_Net
  from ImageLoader import ImageLoader2D


# Checking the number of GPUs available

print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

# Setting the model parameters

img_size = 352
dataset_type = 'etis-laribpolypdb' # Options: kvasir/cvc-clinicdb/cvc-colondb/etis-laribpolypdb
learning_rate = 1e-4
seed_value = 58800
filters = 17 # Number of filters, the paper presents the results with 17 and 34
optimizer = tf.keras.optimizers.RMSprop(learning_rate=learning_rate)

ct = datetime.now()

model_type = "DuckNet"

progress_path = 'ProgressFull/' + dataset_type + '_progress_csv_' + model_type + '_filters_' + str(filters) +  '_' + str(ct) + '.csv'
progressfull_path = 'ProgressFull/' + dataset_type + '_progress_' + model_type + '_filters_' + str(filters) + '_' + str(ct) + '.txt'
plot_path = 'ProgressFull/' + dataset_type + '_progress_plot_' + model_type + '_filters_' + str(filters) + '_' + str(ct) + '.png'
model_path = 'ModelSaveTensorFlow/' + dataset_type + '/' + model_type + '_filters_' + str(filters) + '_' + str(ct)

EPOCHS = 600
min_loss_for_saving = 0.2


# Loading the data

X, Y = ImageLoader2D.load_data(img_size, img_size, -1, 'etis-laribpolypdb')

# Splitting the data, seed for reproducibility

x_train, x_test, y_train, y_test = train_test_split(X, Y, test_size=0.1, shuffle= True, random_state = seed_value)
x_train, x_valid, y_train, y_valid = train_test_split(x_train, y_train, test_size=0.111, shuffle= True, random_state = seed_value)

# Defining the augmentations

aug_train = albu.Compose([
    albu.HorizontalFlip(),
    albu.VerticalFlip(),
    albu.ColorJitter(brightness=(0.6,1.6), contrast=0.2, saturation=0.1, hue=0.01, always_apply=True),
    albu.Affine(scale=(0.5,1.5), translate_percent=(-0.125,0.125), rotate=(-180,180), shear=(-22.5,22), always_apply=True),
])

def augment_images():
    x_train_out = []
    y_train_out = []

    for i in range (len(x_train)):
        ug = aug_train(image=x_train[i], mask=y_train[i])
        x_train_out.append(ug['image'])  
        y_train_out.append(ug['mask'])

    return np.array(x_train_out), np.array(y_train_out)


# Creating the model

gpus = tf.config.list_physical_devices('GPU')
if gpus:
  # Restrict TensorFlow to only allocate 40GB of memory on the GPU
  try:
    tf.config.set_logical_device_configuration(
        gpus[0],
        [tf.config.LogicalDeviceConfiguration(memory_limit=40000)])
    logical_gpus = tf.config.list_logical_devices('GPU')
    print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
  except RuntimeError as e:
    # Virtual devices must be set before GPUs have been initialized
    print(e)


model = DUCK_Net.create_model(img_height=img_size, img_width=img_size, input_chanels=3, out_classes=1, starting_filters=filters)

# Compiling the model

model.compile(optimizer=optimizer, loss=dice_metric_loss)

# Training the model

step = 0

for epoch in range(0, EPOCHS):
    
    print(f'Training, epoch {epoch}')
    print('Learning Rate: ' + str(learning_rate))

    step += 1
        
    image_augmented, mask_augmented = augment_images()
    
    csv_logger = CSVLogger(progress_path, append=True, separator=';')
    
    model.fit(x=image_augmented, y=mask_augmented, epochs=1, batch_size=4, validation_data=(x_valid, y_valid), verbose=1, callbacks=[csv_logger])
    
    prediction_valid = model.predict(x_valid, verbose=0)
    loss_valid = dice_metric_loss(y_valid, prediction_valid)
    
    loss_valid = loss_valid.numpy()
    print("Loss Validation: " + str(loss_valid))
        
    prediction_test = model.predict(x_test, verbose=0)
    loss_test = dice_metric_loss(y_test, prediction_test)
    loss_test = loss_test.numpy()
    print("Loss Test: " + str(loss_test))
        
    with open(progressfull_path, 'a') as f:
        f.write('epoch: ' + str(epoch) + '\nval_loss: ' + str(loss_valid) + '\ntest_loss: ' + str(loss_test) + '\n\n\n')
    
    if min_loss_for_saving > loss_valid:
        min_loss_for_saving = loss_valid
        print("Saved model with val_loss: ", loss_valid)
        model.save(model_path)
        
    del image_augmented
    del mask_augmented

    gc.collect()


# Computing the metrics and saving the results

print("Loading the model")

model = tf.keras.models.load_model(model_path, custom_objects={'dice_metric_loss':dice_metric_loss})

prediction_train = model.predict(x_train, batch_size=4)
prediction_valid = model.predict(x_valid, batch_size=4)
prediction_test = model.predict(x_test, batch_size=4)

print("Predictions done")

dice_train = f1_score(np.ndarray.flatten(np.array(y_train, dtype=bool)),
                           np.ndarray.flatten(prediction_train > 0.5))
dice_test = f1_score(np.ndarray.flatten(np.array(y_test, dtype=bool)),
                          np.ndarray.flatten(prediction_test > 0.5))
dice_valid = f1_score(np.ndarray.flatten(np.array(y_valid, dtype=bool)),
                           np.ndarray.flatten(prediction_valid > 0.5))

print("Dice finished")


miou_train = jaccard_score(np.ndarray.flatten(np.array(y_train, dtype=bool)),
                           np.ndarray.flatten(prediction_train > 0.5))
miou_test = jaccard_score(np.ndarray.flatten(np.array(y_test, dtype=bool)),
                          np.ndarray.flatten(prediction_test > 0.5))
miou_valid = jaccard_score(np.ndarray.flatten(np.array(y_valid, dtype=bool)),
                           np.ndarray.flatten(prediction_valid > 0.5))

print("Miou finished")


precision_train = precision_score(np.ndarray.flatten(np.array(y_train, dtype=bool)),
                                  np.ndarray.flatten(prediction_train > 0.5))
precision_test = precision_score(np.ndarray.flatten(np.array(y_test, dtype=bool)),
                                 np.ndarray.flatten(prediction_test > 0.5))
precision_valid = precision_score(np.ndarray.flatten(np.array(y_valid, dtype=bool)),
                                  np.ndarray.flatten(prediction_valid > 0.5))

print("Precision finished")


recall_train = recall_score(np.ndarray.flatten(np.array(y_train, dtype=bool)),
                            np.ndarray.flatten(prediction_train > 0.5))
recall_test = recall_score(np.ndarray.flatten(np.array(y_test, dtype=bool)),
                           np.ndarray.flatten(prediction_test > 0.5))
recall_valid = recall_score(np.ndarray.flatten(np.array(y_valid, dtype=bool)),
                            np.ndarray.flatten(prediction_valid > 0.5))

print("Recall finished")


accuracy_train = accuracy_score(np.ndarray.flatten(np.array(y_train, dtype=bool)),
                                np.ndarray.flatten(prediction_train > 0.5))
accuracy_test = accuracy_score(np.ndarray.flatten(np.array(y_test, dtype=bool)),
                               np.ndarray.flatten(prediction_test > 0.5))
accuracy_valid = accuracy_score(np.ndarray.flatten(np.array(y_valid, dtype=bool)),
                                np.ndarray.flatten(prediction_valid > 0.5))


print("Accuracy finished")


final_file = 'results_' + model_type + '_' + str(filters) + '_' + dataset_type + '.txt'
print(final_file)

with open(final_file, 'a') as f:
    f.write(dataset_type + '\n\n')
    f.write('dice_train: ' + str(dice_train) + ' dice_valid: ' + str(dice_valid) + ' dice_test: ' + str(dice_test) + '\n\n')
    f.write('miou_train: ' + str(miou_train) + ' miou_valid: ' + str(miou_valid) + ' miou_test: ' + str(miou_test) + '\n\n')
    f.write('precision_train: ' + str(precision_train) + ' precision_valid: ' + str(precision_valid) + ' precision_test: ' + str(precision_test) + '\n\n')
    f.write('recall_train: ' + str(recall_train) + ' recall_valid: ' + str(recall_valid) + ' recall_test: ' + str(recall_test) + '\n\n')
    f.write('accuracy_train: ' + str(accuracy_train) + ' accuracy_valid: ' + str(accuracy_valid) + ' accuracy_test: ' + str(accuracy_test) + '\n\n\n\n')

print('File done')

Test/ Val Loss File: etis-laribpolypdb_progress_DuckNet_filters_17_2024-07-12 11:25:19.256742.txt

Result File: results_DuckNet_17_etis-laribpolypdb.txt

CPU: Intel(R) Xeon(R) Gold 5320 CPU @ 2.20GHz
GPU: NVIDIA A100 80GB