APPLICATION OF TRANSFER LEARNING FOR MEDICAL IMAGE
CLASSIFICATION 1 2 Pak V.O. , Akbarkhujayev S.A. Email: Pak1169@scientifictext.ru
1Pak Vitaliy Olegovich - Master Student;
2Akbarkhujayev Sultonkhuja Akbarkhuja ugli - Master Student, SOFTWARE ENGINEERING FACULTY, TASHKENT UNIVERSITY OF INFORMATION TECHNOLOGIES NAMED AFTER MUHAMMAD AL-KHWARIZMI, TASHKENT, REPUBLIC OF UZBEKISTAN
Abstract: identification of similar types of objects in image processing now become regular task, but when dealing with differing objects it may become complicated. Even it becomes more complex when it is required to not just identify objects, but also categorize them into specific classes. Deep Learning methods proven to excel traditional machine learning techniques in the performance of image processing tasks. One of such methods is Transfer learning which has become prevalent in recent years. This research proposes Transfer Learning method to classify images of gastrointestinal tract (GI) based on Inception-v3 convolution neural network (CNN) model. Keywords: transfer Learning, Inception-v3, Deep Learning, CNN, Kvasir dataset.
ПРИМЕНЕНИЕ ТРАНСФЕРНОГО ОБУЧЕНИЯ ДЛЯ КЛАССИФИКАЦИИ МЕДИЦИНСКИХ ИЗОБРАЖЕНИЙ Пак В.О.1, Акбархужаев С.А.2
1Пак Виталий Олегович - магистрант;
2Акбархужаев Султонхужа Акбархужа угли - магистрант, факультет программной инженерии, Ташкентский университет информационных технологий им. Мухаммада аль-Хорезми, г. Ташкент, Республика Узбекистан
Аннотация: идентификация сходных типов объектов при обработке изображений в настоящее время становится обычной задачей, но при работе с различными объектами она может усложниться. Она становится более сложной, когда требуется не просто идентифицировать объекты, но и классифицировать их по определенным классам. Методы глубокого обучения доказали превосходство над традиционными методами машинного обучения при выполнении задач обработки изображений. Одним из таких методов является трансферное обучение, которое стало распространенным в последние годы. В данном исследовании предлагается метод трансферного обучения для классификации изображений желудочно-кишечного тракта (ЖКТ) на основе модели сверточной нейронной сети Inception-v3 (CNN).
Ключевые слова: трансферное обучение, Inception-v3, глубокое изучение, CNN, набор данных Kvasir.
I. Introduction
In the last decade, several researches were concentrated on deep learning applications for medical images Computer-aided diagnosis (CAD) methods using convolutional neural networks to learn medical image patterns based on a large training dataset [5]. It is reported that deep learning techniques demonstrate more accurate results of image classification comparing to traditional machine learning methods [5]. This research has been focused on gastrointestinal tract medical images. Kvasir dataset [1] has been used. It contains 8000 images of gastrointestinal tract. It represents 8 classes with 1000 images for each class such as dyed lifted polyps, dyed resection margins, esophagitis, normal cecum, normal pylorus, normal z-line, polyps and ulcerative colitis.
II. Research methodology
2.1. Transfer Learning
Transfer Learning is an approach to use gained knowledge while solving one problem to solve different, but related problem [2]. In Deep Learning CNNs trained on ImageNet [3] have been used for classification, clustering or another learning task performed on a new image dataset in order to solve the problem of insufficient data and reduce training time and effort. Transfer Learning has been reported to considerably outperform traditional learning approaches [2]. To conduct Transfer Learning for this research learned features were extracted from a pre-trained CNN, and transferred to a new model. The last layer of the neural network was replaced by a new layer with 8 units representing 8 classes of dataset.
2.2. Inception-v3 Model
Inception-v3 [7] is one of the pretrained models on the TensorFlow open source machine learning platform [9]. It is a new generation of the model after Inception-v1 [8], Inception-v2 [7] in 2015. The Inception-v3 model has been trained on the ImageNet datasets [3], containing the images of 1000 classes. In ImageNet, the error rate of top-5 is 3.5%, the error rate of top-1 dropped to 17.3%.
2.3. Preprocessing
The study was performed in TensorFlow [9]. 70% of data were randomly chosen for training. The remaining 30% were proportioned as 20% and 10% for validation and testing respectively. Bukar and Ugail demonstrated the value of larger training set with the intent to get results of higher accuracy [4]. Inception V3 model requires input images of size 299-by-299-by-3, but the dataset images have different sizes. So the training and test images were resized to height 299 and width 299 before they are input to the pre-training network.
2.4. Evaluation
To assess the performance of built model several metrics have been used such as Accuracy, Precision, Recall. Sokolova and Lapalme (2009) derived these terms from the confusion matrix [6]. Accuracy (1) evaluates the general effectiveness of a classifier; Precision (2) evaluates the correctness of labeling of data labels corresponding to the positive labels given by the classifier; Recall (3) evaluates the correctness of identification of positive labels of a classifier;
Accuracy = —-—— (1) Precision = (2) Recall = (3)
tp + tn+fp+fn tp+fp tp+fn
IILResults
Figure 1 shows the variation of accuracy (left) and loss (right) based on Kvasir dataset. the blue line represents the training set and the orange line represents the validation set.
Training and Validation Accuracy Training and Validation Loss
0.90
035 0.60 075 070 0.65 0.60 055
0 5 10 15 0 5 10 15
Fig. 1. The variation of accuracy and loss on Kvasir dataset
The training accuracy shows the percentage of the images labeled correctly on a batch of images of the training set (higher-better). The validation accuracy is the precision (percentage of correctly-labelled images) on a randomly-selected group of images from the validation set (higher-better). The training and validation losses demonstrate the sum of errors made for each example in training or validation sets (lower - better).
Table 1. Accuracy and Loss values
Index Value
Training Accuracy 0.9%
Validation Accuracy 0.87%
Test Accuracy 0.85%
Training Loss 0.3%
Validation Loss 0.35%
Test Loss 0.38%
Table 2. Precision and Recall values
Classes Precision Recall
dyed lifted polyps 0.72% 0.89%
dyed resection margins 0.89% 0.74%
esophagitis 0.74% 0.81%
normal cecum 0.91% 0.97%
normal pylorus 0.98% 0.95%
normal z-line 0.81% 0.72%
polyps 0.93% 0.80%
ulcerative colitis 0.89% 0.93%
IV. Conclusion
This research proposed a classifier based on the Inception-v3 model of TensorFlow platform, in which the transfer learning technology has been applied to train medical
classification model on kvasir dataset. The classification accuracy of the model is approximately 85% on given dataset. The future work is to study and develop a more effective and accurate model for medical image classification.
References / Список литературы
1. Pogorelov К. et al. "Kvasir: A Multi-Class Image Dataset for Computer Aided Gastrointestinal Disease Detection". In MMSys'17 Proceedings of the 8th ACM on Multimedia Systems Conference (MMSYS). Pages 164-169. Taipei. Taiwan, June 20-23, 2017.
2. Pan S.J., Yang Q. "A survey on transfer learning." IEEE Transactions on knowledge and data engineering. 22 (10). Pages 1345-1359, 2010.
3. Krizhevsky А., Sutskever I. and Hinton Е. "ImageNet Classification with Deep Convolutional Neural Networks." Advances in neural information processing systems, 2012.
4. BukarA.M., UgailH. "Convnet Features for Age Estimation". Bradford. UK. ISBN: 9978-9898533-66-1, 2017.
5. Wang H., Zhou Z., Li Y., Chen Z., Lu Р, Wang W., Liu W. and Yu L. "Comparison of machine learning methods for classifying mediastinal lymph node metastasis of non-small cell lung cancer from 18F-FDG PET/CT images". EJNMMI Research 7:11, DOI 10.1186/s13550-017-0260-9, 2017.
6. Sokolova М. and Lapalme G. "A systematic analysis of performance measures for classification tasks". Inf. Process. Manage, 45 (4):427- 437, July, 2009.
7. Szegedy С., Vanhoucke V., Ioffe S., Shlens J. and Wojna Z. "Rethinking the inception architecture for computer vision". arXiv preprint arXiv:1512.00567, 2015.
8. Christian Szegedy J., Wei Liu et al. "Going Deeper with Convolutions". arXiv:1409.4842,
2014.
9. Abadi Martín et al. "TensorFlow: Large-scale machine learning on heterogeneous systems",
2015.