LiteFeatNet: A parameter-efficient and performance-centric deep learning model for multi-ocular disease identification using intermediate feature reduction from fundus images.

Code ↔ Paper

3 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 3 matches

[1] § 3. Materials and methods › 3.8. Experimental setup ↔ kerasMods.ipynb, lines 559–574 · score 0.68 · ReduceLROnPlateau, TensorFlow, Keras, scheduler, epochs, validation
[2] § 3. Materials and methods › 3.8. Experimental setup ↔ code-nb-p.ipynb, lines 312–316 · score 0.62 · Cross Entropy, learning rate scheduler, Adam, optimizer, loss, train
[3] § 3. Materials and methods › 3.8. Experimental setup ↔ kerasMods.ipynb, lines 559–574 · score 0.53 · ReduceLROnPlateau, Keras, schedulers, trained, model

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The paper is loaded by your browser, with JavaScript: read it at doi.org or on Europe PMC.

The authors' code

Jupyter notebook · 617 lines · 21 KB · MIT · 2 matches

# %%
# libraries
import os
import cv2
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras import backend as K
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.applications import MobileNet,MobileNetV2, VGG16, Xception, NASNetMobile, DenseNet121, EfficientNetB5,EfficientNetB0,EfficientNetV2B0
from tensorflow.keras import Input, Model
from tensorflow.keras.layers import Conv2D, MaxPooling2D,ReLU,SeparableConv2D,Conv2DTranspose,GlobalAveragePooling2D,GlobalMaxPooling2D,Add,Activation, Dense, Dropout, Flatten, BatchNormalization, Concatenate, Activation
from tensorflow.keras.optimizers import Adam, SGD
from tensorflow.keras.regularizers import l2
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import time
import gc
import matplotlib.pyplot as plt
import seaborn as sns
from tensorflow.keras.models import load_model
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, confusion_matrix
# %%
# initial
K.clear_session()
gc.collect()
imgSize = (224,224)
inputShape = tuple(list(imgSize) + [3])
batchSize = 32
print(inputShape)
# %%
# I/O paths
outPath = '/kaggle/working/'
trainImagesPath = '/kaggle/input/dataset/training/images/'
trainLabelsPath = '/kaggle/input/dataset/training/labels.xlsx'
testImagesPath = '/kaggle/input/dataset/testing/images/'
testLabelsPath = '/kaggle/input/dataset/testing/labels.xlsx'
valImagesPath = '/kaggle/input/dataset/validation/images/'
valLabelsPath = '/kaggle/input/dataset/validation/labels.xlsx'
# %%
# model handling
def saveModel(model,savePath,modName):
model.save(savePath + modName + '.h5')
def printModelSummary(model):
print(model.summary())
# %%
# time information
def saveTime(time1,time2,savePath):
time_data = {'Training_Time': [time1],'Testing_Time': [time2]}
time_df = pd.DataFrame(time_data)
time_df.to_excel(savePath + 'time.xlsx',index=False)
# %%
# epch vs. accuracy
def plotAccuracy(hist,savePath):
epochs = range(1, len(hist.history['accuracy']) + 1)
plt.figure(figsize=(8,6))
plt.plot(epochs, hist.history['accuracy'], label='Train Accuracy')
plt.plot(epochs, hist.history['val_accuracy'], label='Validation Accuracy')
plt.xlabel('Epoch', fontsize=12)
plt.ylabel('Accuracy', fontsize=12)
plt.xticks(ticks=range(0, len(epochs)+1, 5))
plt.grid(True,linestyle='--', linewidth=0.5, color='gray')
plt.legend()
plt.savefig(savePath + 'accuracy_plot.png', dpi=1200)
plt.show()
# %%
# epoch vs. loss
def plotLoss(hist,savePath):
epochs = range(1, len(hist.history['loss']) + 1)
plt.figure(figsize=(8,6))
plt.plot(epochs, hist.history['loss'], label='Train Loss')
plt.plot(epochs, hist.history['val_loss'], label='Validation Loss')
plt.xlabel('Epoch',fontsize=12)
plt.ylabel('Loss', fontsize=12)
plt.xticks(ticks=range(0, len(epochs)+1, 5))
plt.grid(True,linestyle='--', linewidth=0.5, color='gray')
plt.legend()
plt.savefig(savePath + 'loss_plot.png', dpi=1200)
plt.show()
# %%
# model history
def saveModelRunHistory(hist,savePath):
epochs = list(range(1, len(hist.history['accuracy']) + 1))
train_acc = hist.history['accuracy']
val_acc = hist.history['val_accuracy']
train_loss = hist.history['loss']
val_loss = hist.history['val_loss']
modelHistory_df = pd.DataFrame({
'Epoch': epochs,
'Training Accuracy': train_acc,
'Validation Accuracy': val_acc,
'Training Loss': train_loss,
'Validation Loss': val_loss
})
modelHistory_df.to_excel(savePath + 'history_file.xlsx', index=False)
# %%
# calculate metrics (overall)
def testMetricsOverall(Ypred,Yact,savePath):
accuracy = accuracy_score(Yact, Ypred)
precision = precision_score(Yact, Ypred, average='weighted', zero_division=0)
recall = recall_score(Yact, Ypred, average='weighted', zero_division=0)
f1 = f1_score(Yact, Ypred, average='weighted', zero_division=0)
print(f"Accuracy: {accuracy}")
print(f"Precision: {precision}")
print(f"Recall: {recall}")
print(f"F1-Score: {f1}")
metrics = {'Metric': ['Accuracy', 'Precision', 'Recall', 'F1-Score'],
'Score': [accuracy, precision, recall, f1]
}
overallMetrics_df = pd.DataFrame(metrics)
overallMetrics_df.to_excel(savePath + 'overall_metrics.xlsx', index=False)
# %%
# calculate metrics (class-wise breakdown)
def testMetricsClassWise(Ypred,Yact,classEncoding,savePath):
classLabels = list(classEncoding.keys())
accuracy_list = []
precision_list = []
recall_list = []
f1_list = []
TP_list = []
TN_list = []
FP_list = []
FN_list = []
TPR_list = []
TNR_list = []
FPR_list = []
FNR_list = []
for label in classLabels:
encoded_label = classEncoding[label]
TP = np.sum((Ypred == encoded_label) & (Yact == encoded_label))
TP_list.append(TP)
FN = np.sum((Ypred != encoded_label) & (Yact == encoded_label))
FN_list.append(FN)
FP = np.sum((Ypred == encoded_label) & (Yact != encoded_label))
FP_list.append(FP)
TN = np.sum((Ypred != encoded_label) & (Yact != encoded_label))
TN_list.append(TN)
accuracy = (TP + TN) / (TP + TN + FP + FN)
accuracy_list.append(accuracy)
precision = precision_score(Yact, Ypred, labels=[encoded_label], average='macro', zero_division=0)
precision_list.append(precision)
recall = recall_score(Yact, Ypred, labels=[encoded_label], average='macro', zero_division=0)
recall_list.append(recall)
f1 = f1_score(Yact, Ypred, labels=[encoded_label], average='macro', zero_division=0)
f1_list.append(f1)
TPR = TP / (TP + FN) if (TP + FN) != 0 else 0
TPR_list.append(TPR)
TNR = TN / (TN + FP) if (TN + FP) != 0 else 0
TNR_list.append(TNR)
FPR = 1 - TNR
FPR_list.append(FPR)
FNR = 1 - TPR
FNR_list.append(FNR)
classwiseMetrics_df = pd.DataFrame({
'Class': classLabels,
'TP':TP_list,
'TN':TN_list,
'FP':FP_list,
'FN':FN_list,
'Accuracy': accuracy_list,
'Precision': precision_list,
'Recall': recall_list,
'F1-Score': f1_list,
'TPR': TPR_list,
'TNR': TNR_list,
'FPR': FPR_list,
'FNR': FNR_list})
classwiseMetrics_df.to_excel(savePath + 'classwise_metrics.xlsx', index=False)
# %%
# create confusion matrix
def createConfusionMatrix(Ypred,Yact,classEncoding,savePath):
classLabels = list(classEncoding.keys())
conf_matrix = confusion_matrix(Yact, Ypred)
plt.figure(figsize=(12, 10))
ax = sns.heatmap(conf_matrix, annot=False, fmt='d', cmap='Blues',cbar=True,
linewidths=0.5, linecolor='gray',
xticklabels=classLabels, yticklabels=classLabels)
for i in range(conf_matrix.shape[0]):
for j in range(conf_matrix.shape[1]):
ax.text(j + 0.5, i + 0.5, str(conf_matrix[i, j]), ha="center", va="center",
color="black", fontsize=12)
cbar = ax.collections[0].colorbar
cbar.ax.yaxis.set_major_formatter(plt.ScalarFormatter())
cbar.ax.yaxis.set_tick_params(labelsize=14)
plt.title('Confusion Matrix', fontsize=16)
plt.xlabel('Predicted Label', fontsize=14)
plt.ylabel('True Label', fontsize=14)
plt.xticks(fontsize=12)
plt.yticks(fontsize=12)
plt.savefig(savePath + 'confusion_matrix.png', dpi=1200)
plt.show()
# %%
# pre-process images
def preProcess(input_img):
output_img = cv2.cvtColor(input_img,cv2.COLOR_BGR2RGB)
output_img = output_img.astype(np.float32) / 255.0
return output_img
# %%
# read images and labels
def readData(imageFolderPath,labelFilePath):
images = []
labels = []
labels_df = pd.read_excel(labelFilePath)
for idx, row in labels_df.iterrows():
img_path = os.path.join(imageFolderPath, row['imageID'])
image = cv2.imread(img_path)
image = cv2.resize(image, imgSize)
image = preProcess(image)
images.append(image)
labels.append(row['disease'])
return np.array(images),np.array(labels)
# %%
def getTotalLabels(trainLabels,testLabels,valLabels):
totalLabels = 0
totalTrainLabels = len(set(list(trainLabels)))
totalTestLabels = len(set(list(testLabels)))
totalValLabels = len(set(list(valLabels)))
if (totalTrainLabels == totalTestLabels and totalTrainLabels == totalValLabels):
totalLabels = totalTrainLabels
else:
print('Total Labels Mismatch in Dataset Splits')
return totalLabels
# %%
def oneHotEncoder(trainLabels, testLabels, valLabels,numLabels):
label_encoder = LabelEncoder()
all_labels = np.concatenate([trainLabels, testLabels, valLabels])
label_encoder.fit(all_labels)
trainLabels_encoded = label_encoder.transform(trainLabels)
testLabels_encoded = label_encoder.transform(testLabels)
valLabels_encoded = label_encoder.transform(valLabels)
class_names = label_encoder.classes_
class_to_encoder = {class_name: idx for idx, class_name in enumerate(class_names)}
trainLabels_1Hot = to_categorical(trainLabels_encoded,num_classes=numLabels)
testLabels_1Hot = to_categorical(testLabels_encoded,num_classes=numLabels)
valLabels_1Hot = to_categorical(valLabels_encoded,num_classes=numLabels)
return trainLabels_1Hot, testLabels_1Hot, valLabels_1Hot,class_to_encoder
# %%
# read images and labels
Xtrain, Ytrain = readData(trainImagesPath,trainLabelsPath)
Xtest, Ytest = readData(testImagesPath,testLabelsPath)
Xval, Yval = readData(valImagesPath,valLabelsPath)
totalLabels = getTotalLabels(Ytrain, Ytest, Yval)
Ytrain, Ytest, Yval, class_encoding = oneHotEncoder(Ytrain, Ytest, Yval,totalLabels)
print(Xtrain.shape)
print(Xtest.shape)
print(Xval.shape)
print(class_encoding)
# %%
# comparison model
def compMod1(inputShape,numLabels):
inp = Input(shape = inputShape)
baseModel = NASNetMobile(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_output = baseModel.output
x = GlobalAveragePooling2D()(baseModel_output)
x = Flatten()(x)
x = Dense(1024,activation='relu')(x)
x = Dropout(0.3)(x)
output = Dense(numLabels,activation='softmax')(x)
model = Model(inp,output,name='compMod1')
opt = Adam(learning_rate=0.001)
model.compile(optimizer = opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# reference model
def ltftnet(inputShape,numLabels):
inp = Input(shape = inputShape)
baseModel = NASNetMobile(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_output = baseModel.get_layer('normal_concat_12').output
x = Conv2D(512,(1,1),activation='relu')(baseModel_output)
x = GlobalAveragePooling2D()(x)
x = BatchNormalization(name='head-batchnorm-1')(x)
x = Dense(512, activation='relu',name='head-dense-2')(x)
x = Dropout(0.3,name='head-dropout-1')(x)
x = Dense(256, activation='relu',name='head-dense-4')(x)
output = Dense(numLabels, activation='softmax',name='class')(x)
model = Model(inp,output,name='usmanrafi')
opt = Adam(learning_rate=1e-4)
model.compile(optimizer = opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# comparison model
def compMod2(inputShape,numLabels):
inp = Input(shape = inputShape)
x = Conv2D(32,(3,3),activation='relu',strides=1)(inp)
x = MaxPooling2D((2,2))(x)
x = Conv2D(64,(3,3),activation='relu',strides=1)(x)
x = MaxPooling2D((2,2))(x)
x = Conv2D(128,(3,3),activation='relu',strides=1)(x)
x = MaxPooling2D((2,2))(x)
x = Conv2D(256,(3,3),activation='relu',strides=1)(x)
x = MaxPooling2D((2,2))(x)
x = Flatten()(x)
x = Dropout(0.4)(x)
output = Dense(numLabels, activation='softmax',name='class')(x)
model = Model(inp,output,name='compMod2')
opt = Adam(learning_rate=0.001)
model.compile(optimizer = opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# comparison model
def compMod3(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel1 = DenseNet121(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel1.trainable = False
baseModel1_output = baseModel1.output
baseModel2 = EfficientNetB5(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel2.trainable = False
baseModel2_output = baseModel2.output
f = Concatenate()([baseModel1_output,baseModel2_output])
gap = GlobalAveragePooling2D()(f)
gmp = GlobalMaxPooling2D()(f)
x = Add()([gap,gmp])
x = Flatten()(x)
x = BatchNormalization()(x)
x = Dropout(0.25)(x)
x = Dense(512,activation='relu')(x)
x = BatchNormalization()(x)
x = Dropout(0.25)(x)
output = Dense(numLabels,activation='softmax')(x)
model = Model(inp,output,name='compMod3')
opt = Adam(learning_rate=0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# comparison model
def transBlock(x):
x = SeparableConv2D(64, (1, 1), padding='same')(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = SeparableConv2D(128, (3, 3), padding='same')(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = SeparableConv2D(256, (1, 1), padding='same')(x)
x = BatchNormalization()(x)
x = ReLU()(x)
x = Conv2DTranspose(256, (3, 3), strides=(2, 2), padding='same')(x)
x = MaxPooling2D((2, 2))(x)
return x
def addLayer(x):
x = SeparableConv2D(256, (1, 1), padding='same')(x)
x = BatchNormalization()(x)
return x
def compMod4(inputShape, numLabels):
inp = Input(shape=inputShape)
vgg_input = Input(shape=inputShape)
vgg_base = VGG16(include_top=False, input_tensor=vgg_input, weights='imagenet')
vgg_base.trainable = False
leftOut = vgg_base.output
leftOut_TransBlock = transBlock(leftOut)
leftOut_Addl = addLayer(leftOut)
leftF = Add()([leftOut_TransBlock, leftOut_Addl])
vgg_model = Model(vgg_input, leftF)
xcep_input = Input(shape=inputShape)
xcep_base = Xception(include_top=False, input_tensor=xcep_input, weights='imagenet')
xcep_base.trainable = False
rightOut = xcep_base.output
rightOut_TransBlock = transBlock(rightOut)
rightOut_Addl = addLayer(rightOut)
rightF = Add()([rightOut_TransBlock, rightOut_Addl])
xcep_model = Model(xcep_input, rightF)
leftF_out = vgg_model(inp)
rightF_out = xcep_model(inp)
f = Add()([leftF_out, rightF_out])
x = Flatten()(f)
x = Dense(256)(x)
x = Dropout(0.2)(x)
x = Dense(128)(x)
x = Dropout(0.3)(x)
x = Dense(64)(x)
output = Dense(numLabels, activation='softmax')(x)
model = Model(inp, output,name='compMod4')
opt = Adam(learning_rate=0.001)
model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod1(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = NASNetMobile(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_secondLastLayer = baseModel.layers[-2].output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod1')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod2(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = MobileNet(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_secondLastLayer = baseModel.layers[-2].output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod2')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod3(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = MobileNetV2(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_secondLastLayer = baseModel.layers[-2].output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod3')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod4(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = DenseNet121(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_secondLastLayer = baseModel.layers[-2].output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod4')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod5(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = EfficientNetB0(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
#baseModel_secondLastLayer = baseModel.layers[-2].output
baseModel_secondLastLayer = baseModel.get_layer('top_dropout').output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod5')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Fine-tuned transfer learning model
def ftmod6(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = EfficientNetV2B0(include_top=True, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
#baseModel_secondLastLayer = baseModel.layers[-2].output
baseModel_secondLastLayer = baseModel.get_layer('top_dropout').output
output = Dense(numLabels,activation='softmax')(baseModel_secondLastLayer)
model = Model(inp,output,name='ftmod6')
opt = Adam(0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# comparison model
def compMod5(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = NASNetMobile(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_output = baseModel.output
x = GlobalAveragePooling2D()(baseModel_output)
output = Dense(numLabels,activation='softmax')(x)
model = Model(inp,output,name="compMod5")
opt = Adam(learning_rate=0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# comparison model
def compMod6(inputShape,numLabels):
inp = Input(shape=inputShape)
baseModel = NASNetMobile(include_top=False, input_shape=inputShape, weights='imagenet',input_tensor=inp)
baseModel.trainable = False
baseModel_output = baseModel.output
x = Flatten()(baseModel_output)
x = Dense(256,activation='relu')(x)
x = Dropout(0.35)(x)
x = Dense(128,activation='relu')(x)
x = Dropout(0.35)(x)
output = Dense(numLabels,activation='softmax')(x)
model = Model(inp,output,name="compMod6")
opt = Adam(learning_rate=0.001)
model.compile(optimizer=opt,loss='categorical_crossentropy',metrics=['accuracy'])
return model
# %%
# Get Model for training
myModel = ltftnet((Xtrain[0].shape),totalLabels)
printModelSummary(myModel)
# %%
# image augmentation
train_datagen = ImageDataGenerator(
rotation_range=30,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest'
)
train_generator = train_datagen.flow(
Xtrain,
Ytrain,
batch_size=batchSize
)
# %%
# LR Scheduler
reduce_lr = tf.keras.callbacks.ReduceLROnPlateau(
monitor='val_accuracy',
factor=0.5,
patience=3,
verbose=1,
min_lr=1e-7
)
# training loop
trainingTime = 0
numEpochs = 30
start = time.time()
history = myModel.fit(train_generator,steps_per_epoch=len(train_generator),epochs=numEpochs,validation_data=(Xval,Yval),shuffle=True,verbose=1,callbacks=[reduce_lr])
end = time.time()
trainingTime = end - start
# %%
# Loading models with saved weights and biases for generalizability
modelPath = 'path/modName.h5'
myModel = load_model(modelPath)
print(myModel.summary())
# %%
# Warmup model
dummy_input = np.random.rand(1, 224, 224, 3).astype(np.float32)
_ = myModel.predict(dummy_input, batch_size=1)
# test
testingTime = 0
start = time.time()
Ypred = myModel.predict(Xtest)
end = time.time()
testingTime = end - start
Ypred = np.argmax(Ypred, axis=1)
Yact = np.argmax(Ytest, axis=1)
# %%
# Print total number of actual labels and model's predicted labels
print(len(Ypred))
print(len(Yact))
# %%
# results
saveTime(trainingTime,testingTime,outPath)
saveModelRunHistory(history,outPath)
saveModel(myModel,outPath,'modelName')
# %%
# results
plotAccuracy(history,outPath)
plotLoss(history,outPath)
# %%
# results
testMetricsOverall(Ypred,Yact,outPath)
testMetricsClassWise(Ypred,Yact,class_encoding,outPath)
createConfusionMatrix(Ypred,Yact,class_encoding,outPath)

kerasMods.ipynb at commit de72ecc, under MIT · at the source

Overview

Authors: Usman Rafi¹, Qamar Nawaz¹, Muhammad Ahsan Latif¹, Aisha Khatoon²

ORCID iDs: Usman Rafi

Institutions: University of Agriculture Faisalabad (Pakistan)

Journal: PloS one, volume 21, issue 5, article e0347867

Dates: received 20 August 2025; accepted 8 April 2026; published online 11 May 2026

Type: Research article · Language: English

License: CC BY

Identifiers: DOI 10.1371/journal.pone.0347867 · PMID 42113782 · PMCID PMC13160359 · OpenAlex W7160855737

Open access: gold, a free copy (OpenAlex)

Status: code verified

Categories: other (modality), human (organism), methods / tools (subfield)

Methods: Machine learning

MeSH: Deep Learning*, Fundus Oculi*, Image Processing, Computer-Assisted*, Retinal Diseases*, Algorithms, Convolutional Neural Networks, Humans (* major topic)

Topic: Retinal Imaging and Analysis (Radiology, Nuclear Medicine and Imaging, Medicine), according to OpenAlex

Citations: not cited yet (Europe PMC); 66 references in the paper

Abstract

Convolutional Neural Networks (CNNs) require a larger amount of input samples and computing resources to learn discriminative features for accurate identification of multiple retinal conditions, making the development and deployment of such models challenging on limited computing resources. This study presents a robust CNN (called LiteFeatNet) that requires fewer trainable parameters, computational resources, and processing time for accurate prediction. To enhance robustness and reduce computational time, a pre-trained NASNetMobile backbone is employed, and a method for time-efficient discriminative feature extraction from deep intermediate layers is proposed. The extracted features are refined using a spatially-aware feature map reduction module and classified using a custom classification module with fewer number of trainable parameters, reduced computational resource requirements, and computational time. Experiments are conducted using 1824 images from three distinct class labels in the Retinal Fundus Multi-Disease Image Dataset (RFMiD), with a 60:20:20 train-validation-test split. The LiteFeatNet architecture has a compact size (19.87 MB) and was trained using a standard pre-processing pipeline and training configurations. It outperformed twelve state-of-the-art models, achieving the highest testing accuracy of 90.33%, precision of 90.69%, recall of 90.33%, and F1-score of 90.27%, with a fast, impressive inference time of 4 milliseconds per image. Further, a generalizability study was also conducted using an external dataset, RFMiD 2.0, and the LiteFeatNet achieved competitive performance with quicker testing time compared to other architectures. To evaluate the scalability and adaptability of our proposed integrated framework for larger multi-class problems, we assessed the scalability by increasing class label complexity using two additional disease categories. Results validated the effectiveness and computational efficiency of this integrated framework compared with 9 baseline architectures. An ablation study was also conducted using the LiteFeatNet and two top-performing transfer learning architectures to validate that the synergistic combination of deep feature extraction and feature map refinement is the primary design decision behind the success of the LiteFeatNet architecture. The evaluation metrics, thus obtained, strongly suggest that the proposed LiteFeatNet is lightweight, fast, and robust, rendering it suitable for deployment in low-resource clinical settings.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repository

Its files are read in the Code ↔ Paper reader above, with 3 matches between paragraphs and lines of code.

usmanrafics/LTFTNET

License: MIT

State: the link answers, verified on 28 September 2026

Evidence: files inventoried

Commit: de72ecc81c4be3bb15f7468cfb0d50aa766d7999, 10 August 2026

Languages: Jupyter (3)

Size: 5 files, 3 scripts

Software Heritage: not archived

Found in: “Data Availability”

Holds: README, license file, 3 notebooks

Not found: CITATION.cff, environment file, tests, continuous integration, documentation

Tools: Matplotlib (3 files), NumPy (3 files), OpenCV (3 files), pandas (3 files), scikit-learn (3 files), seaborn (3 files), Keras (2 files), TensorFlow (2 files), Pillow (1 file), PyTorch (1 file)

Availability: 1 check, the latest on 28 September 2026: the link answers

28 September 2026: the link answers

5 files

code-nb-k.ipynb, Jupyter, 876 lines
code-nb-p.ipynb, Jupyter, 481 lines, 1 match
kerasMods.ipynb, Jupyter, 617 lines, 2 matches
LICENSE, License, 21 lines
README.md, Text, 45 lines

At the source: github.com/usmanrafics/LTFTNET

The paper's code and data availability statement is in the Data section.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
3 scripts, each with its path and the digest of its content;
3 matches between paragraphs of the paper and lines of the code (method lexical-v1);
neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

Datasets cited

data.mendeley.com/datasets/s9bfhswzjb, at Mendeley Data; found in the references
kaggle.com/datasets/andrewmvd, at Kaggle; found in the references
kaggle.com/datasets/linabennaa, at Kaggle; found in the references
zenodo:7505822, at Zenodo; found in the references

Data Availability

All three datasets can be found at the following recommended repositories. Dataset1 - RFMiD 1: https://www.kaggle.com/datasets/andrewmvd/retinal-disease-classification Dataset2 - RFMiD 2.0: https://zenodo.org/records/7505822 Dataset3 - Eye Disease Dataset https://www.kaggle.com/datasets/linabennaa/eye-disease-image-dataset-mendeley The code developed for this study is made publicly available on GitHub. The code can be accessed at the following URL: https://github.com/usmanrafics/LTFTNET.

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 28 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 4 authors, 7 MeSH terms, 47 references.

Cite

This paper

Rafi, U., Nawaz, Q., Latif, M. A., & Khatoon, A. (2026). LiteFeatNet: A parameter-efficient and performance-centric deep learning model for multi-ocular disease identification using intermediate feature reduction from fundus images. PloS one, 21(5), e0347867. https://doi.org/10.1371/journal.pone.0347867

BibTeX

@article{rafi2026litefeatnet, author = {Rafi, Usman and Nawaz, Qamar and Latif, Muhammad Ahsan and Khatoon, Aisha}, title = {{LiteFeatNet: A parameter-efficient and performance-centric deep learning model for multi-ocular disease identification using intermediate feature reduction from fundus images}}, journal = {PloS one}, year = {2026}, month = may, volume = {21}, number = {5}, pages = {e0347867}, publisher = {PLOS}, issn = {1932-6203}, doi = {10.1371/journal.pone.0347867}, url = {https://doi.org/10.1371/journal.pone.0347867}, pmid = {42113782}, pmcid = {PMC13160359} }

RIS

TY - JOUR AU - Rafi, Usman AU - Nawaz, Qamar AU - Latif, Muhammad Ahsan AU - Khatoon, Aisha TI - LiteFeatNet: A parameter-efficient and performance-centric deep learning model for multi-ocular disease identification using intermediate feature reduction from fundus images T2 - PloS one J2 - PLoS One PY - 2026 DA - 2026/05/11 VL - 21 IS - 5 SP - e0347867 SN - 1932-6203 PB - PLOS DO - 10.1371/journal.pone.0347867 UR - https://doi.org/10.1371/journal.pone.0347867 LA - en ER -

CSL-JSON

{ "id": "10.1371/journal.pone.0347867", "type": "article-journal", "title": "LiteFeatNet: A parameter-efficient and performance-centric deep learning model for multi-ocular disease identification using intermediate feature reduction from fundus images", "container-title": "PloS one", "author": [ { "family": "Rafi", "given": "Usman" }, { "family": "Nawaz", "given": "Qamar" }, { "family": "Latif", "given": "Muhammad Ahsan" }, { "family": "Khatoon", "given": "Aisha" } ], "container-title-short": "PLoS One", "volume": "21", "issue": "5", "page": "e0347867", "DOI": "10.1371/journal.pone.0347867", "PMID": "42113782", "PMCID": "PMC13160359", "ISSN": "1932-6203", "publisher": "PLOS", "URL": "https://doi.org/10.1371/journal.pone.0347867", "language": "en", "issued": { "date-parts": [ [ 2026, 5, 11 ] ] } }

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1038/s41597-025-05174-7 [code]: A large-scale MEG and EEG dataset for object recognition in naturalistic scenes
Journal: n/a
In common: Keras, TensorFlow, OpenCV, 7 other tools, methods / tools
[2] doi:10.1126/sciadv.aed3650 [code]: Truthful visualizations for mass spectrometry imaging enable high-spatial-resolution interactive <i>m/z</i> mapping and exploration.
Journal: Science advances
In common: Keras, TensorFlow, OpenCV, 6 other tools, other, methods / tools
[3] doi:10.1038/s42003-026-10957-8 [code]: Brain defence by the extracellular matrix protein Cochlin.
Journal: Communications biology
In common: Keras, TensorFlow, OpenCV, 7 other tools
[4] doi:10.1371/journal.pcbi.1014571 [code]: SynAPSeg: A novel dataset and image analysis framework for deep learning-based synapse detection and quantification.
Journal: PLoS computational biology
In common: Keras, TensorFlow, OpenCV, 7 other tools
[5] doi:10.1364/boe.605322 [code]: Generalized plaque digitization framework for multi-dimensional mesoscopic images.
Journal: Biomedical optics express
In common: Keras, TensorFlow, OpenCV, 7 other tools
[6] doi:10.1038/s41467-026-72057-9 [code]: Sex-specific behavioral feedback modulates sensorimotor processing and drives flexible social behavior.
Journal: Nature communications
In common: Keras, TensorFlow, OpenCV, 7 other tools
[7] doi:10.1038/s41598-026-48475-6 [code]: Optimized U-net model for precise retinal blood vessel segmentation from colour fundus images.
Journal: Scientific reports
In common: Keras, TensorFlow, OpenCV, 5 other tools, other, methods / tools
[8] doi:10.1038/s41592-026-03057-2 [code]: CREsted: modeling genomic and synthetic cell-type-specific enhancers across tissues and species.
Journal: Nature methods
In common: Keras, TensorFlow, Pillow, 6 other tools, methods / tools
[9] doi:10.1126/sciadv.aee6952 [code]: Wafer-scale SOT-MRAM for analog crossbar array applications.
Journal: Science advances
In common: Keras, TensorFlow, OpenCV, 6 other tools
[10] doi:10.1093/jnen/nlaf152 [code]: Clinical and pathologic correlations of machine learning quantification of Aβ deposits across 3 brain regions of decedents with Alzheimer disease.
Journal: Journal of neuropathology and experimental neurology
In common: Keras, TensorFlow, OpenCV, 6 other tools

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.