This is to confirm that CHAN XIAO JING (ID No: 1804918) has completed his final year project titled “OPTIMAL SUBCUTANEOUS VENOUS NEAR INFRARED EXTRACTION USING DEEP LEARNING” under the supervision of DR. GOH CHUAN MENG (Supervisor) from the Department of COMPUTING, Faculty of INFORMATION AND COMMUNICATION TECHNOLOGIES and TS DR. TAN HUNG KHOON (moderator) from the Computer Science Department of the Faculty of Information and Communication Technologies.
I certify that this report entitled “AN OPTIMAL NEAR-INFRARED SUBCUTANEOUS VEIN EXTRACTION USING DEEP LEARNING” is my own work, except as cited in the references. Therefore, this project has proposed an optimal near-infrared subcutaneous vein extraction technique using deep learning.
LIST OF TABLES
LIST OF ABBREVIATIONS
Introduction
- Problem Statement and Motivation
- To assess the model capability on extracting veins from higher-difficulty forearm images
- To assess the performance of subcutaneous vein segmentation using unsupervised learning technique
- Contributions
- Background Information .1 Intravenous (IV) Access
The goal was to increase the ability of the proposed model to extract veins from different features of the forearm, as well as to evaluate the performance of the model when the outline of the vein was not so visible on the patient's forearm. Thus, unsupervised learning was implemented in the model evaluation phase by evaluating control points selected from random directions of forearm veins for the calculation of performance metrics. Deep learning is a subset of the machine learning family that can be implemented using artificial neural networks.
The bottleneck layer mediates between the contracting path and the expanding path by using normal convolutional layers followed by an up-convolutional layer. The pooling layer for Figure 1.4 Comparison between classical neural network and deep learning neural network.
Literature Review
- Historical Development of Subcutaneous Vein Feature Extraction
- Feature Extraction Based on Deep Learning
- Feature Extraction using CNN
- Feature Extraction using DenseNet
- Feature Extraction using AlexNet
- Feature Extraction using VGG-Net
- Feature Extraction using ResNet
- Feature Extraction using Inception-ResNet
- Feature Extraction using U-Net
The following section discusses several deep learning techniques used in vein feature extraction. In addition, Wang and Qin [6] implemented an end-to-end U-Net model for vein feature extraction. Author (year) Technique used Advantages Disadvantages. A) Extraction of vein features using CNN Qin and El-Yacoubi.
30] A combined selective network consisting of two-channel and two-stream CNNs for vein feature extraction. Less effective in the image preprocessing and parameter adjustment stages, which may affect the vein feature extraction process.
System Model and Design
- Design Specifications
- Methodologies and General Work Procedures
- Tools to Use
- User Requirements
- Initial Optimized U-Net Model Architecture
- System Design
- Data Augmentation
- Phase 2 Model Training
- Model Prediction on Unsupervised Test Set
- Hypermodel Performance Evaluation
- Implementation Issues and Challenges
- Hardware resource limitation
A confusion matrix would be calculated to represent the values of four measurements, as shown in the colored table cells in Table 3.2. In the first optimization, the model consisted of four shrink blocks, the bottleneck block and four expansion blocks. In the optimized model, each expansion block contained a transposed convolution layer, a concatenated layer, a dropout layer with a dropout ratio of 0.1, and two 3x3 convolution layers.
The output layer was a 1x1 convolution layer with the application of sigmoid activation instead of softmax activation in the original U-Net architecture. The second optimization was performed on top of the first optimized model by adding batch normalization layers to the contracting path. Thus, to reduce the risk of data overfitting, the complexity of the U-Net model was reduced by using the highest filter size in the bottleneck layers as 512 instead of .
In the output layer, which is a 1×1 convolutional layer, the softmax activation in the original U-Net architecture has been replaced by a sigmoid activation. In the first phase, the U-Net model was further optimized by reducing the number of hidden layers. During model training, all hyperparameters used in the .fit command, which were epochs, steps per epoch, validation steps, and also the data augmentation generator, were similar to those specified from phase 1 model training.
In the hyperparameter fine-tuning phase, the Keras Tuner library was used to determine the optimal hyperparameter defined in the model. In the model prediction phase, an unsupervised dataset with no ground truth collected by a research group from Universiti Tunku Abdul Rahman, with some sample images shown in Figure 3.17, was used to investigate unsupervised subcutaneous vein segmentation using the proposed lightweight U-Net model. When the mouse was clicked on a particular point, a coordinate was displayed on the image as shown in Figure 3-20, and added to an array for later use.
Therefore, this project tried to overcome this limitation by implementing data augmentation in the training images to generate large groups of images from the existing images.
Experiments and System Evaluation
- Phase 1 Model Training
- Analysis of Model Training on Optimized Network Architecture
- Analysis of Model Training on Data Augmentation
- Analysis of Model Training on Different Hyperparameters
- Model Training on Different Learning Rates
- Model Training on Different Activation Functions
- Hyperparameter Fine-Tuning
- Model Evaluation on Unsupervised Test Set Images
This was because the cube coefficient measured the number of positive pixels that can be determined by the model to detect similarity between predicted outputs and the ground truths. The Dice coefficient scores and the predicted output were shown in Table 4.1 and Figure 4.3 respectively. The low cube coefficient score in Table 4.1 indicated that this optimized model was less effective at correctly identifying positive pixels, resulting in less overlap of vein pixels between predicted output and ground truth.
Batch normalization can preserve the input with variance of Table 4.1 Dice coefficient scores for the first optimized network architecture. Figure 4.4 showed less noise in the predicted output, especially for Table 4.2 Dice coefficient scores for the second optimized network architecture. 66 Based on the result obtained, the second combination achieved the highest dice coefficient of all combinations, followed by the first combination.
From the results shown above, it was observed that the model using a learning rate of 0.0001 achieved the highest cube coefficient score and clearest predicted output among all the learning rates. This experiment had used 10, 15 and 25 epochs to further reduce cube coefficient loss and to observe underfitting or overfitting in the model. However, the first lightweight model had a slightly lower validation cube coefficient than training cube coefficient.
79 Figure 4.14 showed the final cube coefficient scores obtained from hypermodel retraining, which were 0.8513 for training cube coefficient and 0.8470 for validation cube coefficient. Both cube coefficient scores were higher than the cube coefficients obtained in the original lightweight model (training cube coefficient = 0.8338, validation cube coefficient, indicating that the optimal hyperparameters had indeed improved the performance of the hypermodel effectively. 81 Based on the results shown above, the hypermodel had obtained a higher cube coefficient compared to the cube coefficient scores obtained in model training and rehabilitation phases.
Conclusion and Future Work
- Conclusion and Objectives Evaluation
- Novelties and Contributions
- Future Work
84 and the accuracy (0.9918) achieved, showing that the proposed model can extract and locate the subcutaneous veins accurately and effectively based on the observation made from the selected control points. Third, the unsupervised learning technique was applied to evaluate the performance of the proposed model in locating and extracting real subcutaneous veins. Therefore, it can be concluded that the proposed optimized lightweight U-Net model had improved the ability and performance of current venipuncture machines in performing intravenous procedures and would reduce the error of wrong vein detection.
The proposed model could be used in venipuncture devices to locate subcutaneous veins for IV procedures. The main contribution of this project was the implementation of unsupervised learning concepts to evaluate model performance. As found in the literature review and other existing works, most vein extraction researches for biometric verification and identity authentication have used ground-truth datasets to perform experiments.
There was a lack of forearm vein datasets available, not to mention forearm vein datasets with ground truth, because most of the vein datasets were composed of palmar or finger veins. To overcome the issue of limited biomedical data, the proposed model was first trained on a supervised soil dataset and then evaluated using an unsupervised soil dataset. For each unsupervised forearm images, 20 control points were selected from random directions to evaluate the extraction outcomes of true and false vein pixels by comparing the values of the control points between the unsupervised forearm images and the predicted outputs.
This research project had selected control points manually from the unsupervised forearm images, which was a cumbersome and time-consuming process throughout the project pipeline. In this context, a larger pool of datasets with the corresponding ground truth would be needed to further validate the robustness of the proposed optimized lightweight U-Net model to extract subcutaneous vein accurately. Apart from that, improvement will be sought by conducting more studies to improve the current U-Net architecture for a more sophisticated and prominent deep learning based vein extraction procedure.
Valli, “Securing deep learning-based edge finger vein biometrics with binary decision diagram,” IEEE Trans. El-Yacoubi, “Deep representation-based feature extraction and restoration for finger-vein authentication,” IEEE Trans. Kang, “A novel finger vein verification system based on two-stream convolutional network learning,” Neurocomputing , vol.
Rosdi, “Identification of finger veins using deep coupled convolutional neural network,” Journal of King Saud University – Computer and Information Sciences, vol. Campisi, “Real-time biometric finger vein recognition using deep neural networks,” IEEE Trans. Park, “Densely connected convolutional network-based finger vein recognition using score-level fusion with shape and texture images,” IEEE Access , vol.
Chen, “Finger vein recognition based on deep learning,” presented at the 12th IEEE Conference on Industrial Electronics and Applications (ICIEA), Siem Reap, Cambodia, June Zhang, “Finger vein recognition based on convolutional neural network", presented at the 2017 International Conference on Electronic Information Technology and Computer Engineering (EITCE 2017), Zhuhai, China, September Elsheikh, "Finger vein identification based on AlexNet transfer learning", presented at the 7th Conference of the International Conference on Computer and Communication Engineering (ICCCE) 2018), Kuala Lumpur, Malaysia, September.
Park, “Convolutional neural network-based finger vein recognition using NIR image sensors,” Sensors (Basel), vol. Uhl, "Finger vein recognition by improved CNN segmentation from joint training with automatically and manually generated labels". Piuri, “Finger Vein Verification Algorithm Based on Fully Convolutional Neural Network and Conditional Random Field,” IEEE Access, vol.
APPENDIX A
Weekly Report
FINAL YEAR PROJECT WEEKLY REPORT
- WORK DONE
- SELF EVALUATION OF THE PROGRESS
- WORK TO BE DONE
- PROBLEMS ENCOUNTERED
Need to read more journals and articles related to subcutaneous vein segmentation using U-Net as the baseline model. Need to speed up report writing progress as next week is draft submission.
PLAGIARISM CHECK RESULT
CHECKLIST
UNIVERSITI TUNKU ABDUL RAHMAN FACULTY OF INFORMATION & COMMUNICATION
TECHNOLOGY (KAMPAR CAMPUS)