Abstract
Drowning death rate is high in Japan and its diagnosis is still one of the most challenging tasks in the field of forensics due to the complex interpretation of its pathology. Postmortem lungs computed tomography (CT) images can be used for interpretation of forensic pathology due to its benefits but shortage of specialists is a critical problem. Also, manually interpreting CT images is a tiring and time-taking process. In this paper, we proposed a computer-aided diagnosis system based on a deep convolutional neural network (DCNN) for classifying the post-mortem lungs CT images into drowning and non-drowning. A pre-trained DCNN was implemented in this study for classification of post-mortem lungs CT images. The DCNN was trained and tested using a post-mortem lungs CT image database obtained from Tohoku University Autopsy Imaging Center. The training process involves fine-tuning. The experimental results demonstrated a receiver operating characteristic (ROC) curve and an area under the curve (AUC) of 95 percent was achieved in drowning detection using the post-mortem lungs CT images.
| Original language | English |
|---|---|
| Title of host publication | Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021 |
| Editors | Yufei Huang, Lukasz Kurgan, Feng Luo, Xiaohua Tony Hu, Yidong Chen, Edward Dougherty, Andrzej Kloczkowski, Yaohang Li |
| Publisher | Institute of Electrical and Electronics Engineers Inc. |
| Pages | 2309-2313 |
| Number of pages | 5 |
| ISBN (Electronic) | 9781665401265 |
| DOIs | |
| Publication status | Published - 2021 |
| Event | 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021 - Virtual, Online, United States Duration: 2021 Dec 9 → 2021 Dec 12 |
Publication series
| Name | Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021 |
|---|
Conference
| Conference | 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021 |
|---|---|
| Country/Territory | United States |
| City | Virtual, Online |
| Period | 21/12/9 → 21/12/12 |
Keywords
- computer-aided diagnosis
- deep convolutional neural network
- drowning
- post-mortem CT
ASJC Scopus subject areas
- Artificial Intelligence
- Computer Science Applications
- Biomedical Engineering
- Health Informatics
- Information Systems and Management
Access to Document
Other files and links
Fingerprint
Dive into the research topics of 'Deep CNN-Based Computer-Aided Diagnosis for Drowning Detection using Post-mortem Lungs CT Images'. Together they form a unique fingerprint.Cite this
- APA
- Standard
- Harvard
- Vancouver
- Author
- BIBTEX
- RIS
Deep CNN-Based Computer-Aided Diagnosis for Drowning Detection using Post-mortem Lungs CT Images. / Qureshi, Amber Habib; Zhang, Xiaoyong; Ichiji, Kei et al.
Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021. ed. / Yufei Huang; Lukasz Kurgan; Feng Luo; Xiaohua Tony Hu; Yidong Chen; Edward Dougherty; Andrzej Kloczkowski; Yaohang Li. Institute of Electrical and Electronics Engineers Inc., 2021. p. 2309-2313 (Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution
}
TY - GEN
T1 - Deep CNN-Based Computer-Aided Diagnosis for Drowning Detection using Post-mortem Lungs CT Images
AU - Qureshi, Amber Habib
AU - Zhang, Xiaoyong
AU - Ichiji, Kei
AU - Kawasumi, Yusuke
AU - Usui, Akihito
AU - Funayama, Masato
AU - Homma, Noriyasu
N1 - Funding Information: Fig. 2. Network structure of VGG16 [11]. C. Experimental method In general, deep learning requires a huge amount of data, but it is difficult to obtain such large number of post-mortem lungs CT images. In order to solve this problem, we used a transfer learning technique in our experiment. Transfer learning is a method by which we can apply information that has been learned previously from a DL model and can be applied to solve new problems [12]. Training any DCNN model from scratch needs a lot of data because they have millions of trainable parameters. In such cases small datasets would not be enough to get good generalization of the model. Models like ResNet, VGG16, Xception or MobileNet etc. can be reused using their pre-trained weights and this is called transfer learning. All these models are pre-trained on the ImageNet [13] dataset that contain 1.4 million natural images of 1,000 classes. It has been reported that the earlier layers of the DCNN trained with the ImageNet are capable of extracting the generic image features, such as edge, colour blob, and that these earlier layers can be transfer to other classification tasks [14]. In this experiment, transfer learning of a pre-trained VGG16 has been used for training the network that can classify post-mortem lungs CT images into drowning and non-drowning. Experiment was carried out using python language with open library Tensorflow and Keras on a GPU (NVIDIA GeForce RTX 2080 Ti) with the CUDA Library. In addition, to solve the overfitting problem, in this experiment, we used data augmentation technique. Fig. 3 shows the post-mortem lungs CT images after augmentation. Data augmentation geometrically transforms the images using a range of operations such as shifts, flips, zooms and much more. Overfitting is caused by having too few samples to learn from. Data augmentation is a powerful way to fight overfitting. Fine-tuning is a widely used technique for model reuse. It will speed up the training process, gives better performance as compared to training from scratch [15] and also overcome the small dataset problem. Fine-tuning involves unfreezing some of the layers in a frozen convolution base network of any CNN and trained these unfreeze layers with the fully connected classifier. The steps for fine-tuning VGG16 are as follow: 1. We added our own fully connected classifier after an already-trained VGG16 base network e.g. in fig. 4. Base network consist of convolution block 1 to 5. 2. After that we freeze the convolution block 1 to 3 of a base network (to keep the weights untrainable), and unfreeze convolution block 4 and 5 (to keep the weights trainable) in Fig. 4, and started training our own fully connected classifier with the unfreeze convolution block 4 and 5. Overall fine-tuned VGG16 is shown in Fig.4. It will take the augmented post-mortem lungs CT images and will process it through several layers i.e. convolutional, max-pooling, flatten, dense and drop out layers and then the output will be generated as drowning or non-drowning. Lastly, we evaluated our model on the test and validation data and checked the accuracy of our system with and without data augmentation. We used binary cross entropy as a cost function and Adam as an optimizer. Hyper parameters were chosen experimentally, e.g. 2×10-5 learning rate, 200 epochs and 20 batch size. As performance indices, a receiver operating characteristic (ROC) curve and an area under the curve (AUC) are calculated from the true positive ratio (TPR) and the false positive rate (FPR). III. RESULTS AND DISCUSSION Fig. 5 represents the training and validation loss for VGG16 model with and without data augmentation. The values for training and validation loss is lower when doing data augmentation. The above mentioned graph clearly indicates that by doing fine-tuning and data augmentation we can reduce both the losses as well. The more precise analysis of the proposed method can be observed from receiver operating characteristics (ROC) curve and area under the curve (AUC) which were calculated from the true positive ratio (TPR) and the false positive ratio (FPR). The ROC curves are shown in Fig. 6. The AUC = 88% was obtained without data augmentation and AUC = 95% was obtained with data augmentation for post-mortem lungs CT image based classification. Fig. 3. Example of geometrically transformed post-mortem lungs CT images after data augmentation. Fig. 4. Structure of the fine-tuned VGG16 Fig. 5. Loss graph for VGG16 showing training and validation loss is lower with augmentation. Fig. 7 (a) shows an example of true positive (TP) that is a drowning case, correctly classified by the proposed system. The ground glass opacity (GGO) can be seen as white or light gray mosaic pattern in the lung area. It is mainly because of water in the lungs. In fact, the GGO is one of the common findings in drowning cases. Fig. 7 (b) shows an example of false positive (FP) that is actually a non-drowning case, but miss-classified by the proposed DCNN. This is a typical false positive case and difficult to be classified correctly because of the presence of whitish GGO. Fig. 7 (c) shows an example of true negative (TN) that is a non-drowning case, correctly classified by the proposed DCNN. In contrast to the GGO of positive cases in Figs. 7 (a) and (b), the lung area is clear and seen as mostly black because of air instead of abnormal water in the lung. Fig. 7 (d) shows an example false negative (FN) that is an apical lung image of a drowning case, but miss-classified by the DCNN. It is a difficult example to be classified because the lung area is small and not much information or significant feature can be seen in the image. IV. CONCLUSIONS In this study, we proposed a CAD system based on VGG16 for drowning diagnosis using post-mortem lungs CT images. Experimental results demonstrated that the proposed system can well perform the drowning classification task as compared to AlexNet. We also compared VGG16 results trained with and without data augmentation. The experimental results demonstrated an AUC-ROC of 88% for pre-trained VGG16 without data augmentation and 95% for pre-trained VGG16 with data augmentation. More comprehensive analysis and comparison of the DCNN performance is under preparation to address the black box problem of deep learning and to further improve the performance. Fig. 6. ROC curves with and without data augmentation. (a) (b) (c) (d) Fig. 7. Examples of the post-mortem lungs CT images used in the experiment. (a) A TP case with GGO findings. (b) A FP case with GGO findings. (c) A TN case with clear lung. (d) A FN case of apical lungs. ACKNOWLEDGMENT This work was partially supported by Autopsy imaging Center, Tohoku University Graduate School of Medicine, and JSPS KAKENHI Grant Numbers are JP18K19892, and JP19H04479. REFERENCES [1] D. Meddings, A. A. Hyder, J. Ozanne-Smith & A Rahman, "Global Report on Drowning: Preventing a Leading Killer," World Health Organization, 2014, pp. 76. http://www.who.int/violence_injury_prevention/global_report_drowning/en/ [2] Ministry of Health, Labor and Welfare, "Vital Statistics,", 2018. [3] A. Farrugia, B. Ludes, "Diagnostic of Drowning in Forensic Medicine," Forensic Medicine - From Old Problems to New Challenges, 2011, D. N. Vieira Ed. Intech. [4] M. Baglivo, S. Winklhofer, G. M. Hatch, G. Ampanozi, M. J. Thali, T. D. Ruder, "The rise of forensic and post-mortem radiology—analysis of the literature between the year 2000 and 2011," Journal of Forensic Radiology and Imaging, 2013, vol. 1, pp. 3-9. [5] M. H. A. Piette, E. A. De Letter, "Drowning: Still a difficult autopsy diagnosis," Forensic Science International, 2006, vol. 163, pp. 1-9. [6] A. Usui, Y. Kawasumi, M. Funayama, H. Saito, "Post-mortem lung features in drowning cases on computed tomography," Japanese Journal of Radiology, 2014, vol. 32, pp. 414-420. [7] A. Christe, E. Aghayev, C. Jackowski, M.J. Thali, P. Vock, "Drowning-post-mortem imaging findings by computed tomography," European Journal of Radiology, 2008, vol. 18, pp. 283-290. [8] Y. Kawasumi, T. Kawabata, Y. Sugai, A. Usui, Y. Hosokai, M. Sato, H. Saito, "Diagnosis of drowning using post-mortem computed tomography based on the volume and density of fluid accumulation in the maxillary and sphenoid sinuses," European Journal of Radiology, 2013, vol. 82, pp. e562-e566. [9] N. Homma, Xiaoyong Zhang, Amber Qureshi, "A Deep Learning Aided Drowning Diagnosis for Forensic Investigations using Post-Mortem Lung CT Images," 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 2020, pp. 1262-1265. [10] K. Simonyan, A. Zisserman, "Very deep convolutional networks for large-scale image recognition," 2015 International Conference on Learning Representations, 2015, pp. 1409-1556. [11] D. Sarkar, "A Comprehensive Hands-on Guide to Transfer Learning with Real-World Applications in Deep Learning," Towards Data Science Web, 2018. [12] M. Oquab, L. Bottou, I. Laptev and J. Sivic, "Learning and Transferring Mid-level Image Representations Using Convolutional Neural Networks," 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1717-1724. [13] J. Deng, W. Dong, R. Socher, L. J. Li, K. Li and L. Fei-Fei, "ImageNet: A Large-Scale Hierarchical Image Database," 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248-255. [14] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng and T. Darrell, "DeCAF: A deep convolutional activation feature for generic visual recognition," Proceedings of the 31st International Conference on Machine Learning, 2014, vol. 32, pp. 647-655. [15] Shallu, M. Rajesh, "Breast cancer histology images classification: Training from scratch or transfer learning?," ICT Express, 2018, vol. 4, pp. 247-254. Publisher Copyright: © 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - Drowning death rate is high in Japan and its diagnosis is still one of the most challenging tasks in the field of forensics due to the complex interpretation of its pathology. Postmortem lungs computed tomography (CT) images can be used for interpretation of forensic pathology due to its benefits but shortage of specialists is a critical problem. Also, manually interpreting CT images is a tiring and time-taking process. In this paper, we proposed a computer-aided diagnosis system based on a deep convolutional neural network (DCNN) for classifying the post-mortem lungs CT images into drowning and non-drowning. A pre-trained DCNN was implemented in this study for classification of post-mortem lungs CT images. The DCNN was trained and tested using a post-mortem lungs CT image database obtained from Tohoku University Autopsy Imaging Center. The training process involves fine-tuning. The experimental results demonstrated a receiver operating characteristic (ROC) curve and an area under the curve (AUC) of 95 percent was achieved in drowning detection using the post-mortem lungs CT images.
AB - Drowning death rate is high in Japan and its diagnosis is still one of the most challenging tasks in the field of forensics due to the complex interpretation of its pathology. Postmortem lungs computed tomography (CT) images can be used for interpretation of forensic pathology due to its benefits but shortage of specialists is a critical problem. Also, manually interpreting CT images is a tiring and time-taking process. In this paper, we proposed a computer-aided diagnosis system based on a deep convolutional neural network (DCNN) for classifying the post-mortem lungs CT images into drowning and non-drowning. A pre-trained DCNN was implemented in this study for classification of post-mortem lungs CT images. The DCNN was trained and tested using a post-mortem lungs CT image database obtained from Tohoku University Autopsy Imaging Center. The training process involves fine-tuning. The experimental results demonstrated a receiver operating characteristic (ROC) curve and an area under the curve (AUC) of 95 percent was achieved in drowning detection using the post-mortem lungs CT images.
KW - computer-aided diagnosis
KW - deep convolutional neural network
KW - drowning
KW - post-mortem CT
UR - http://www.scopus.com/inward/record.url?scp=85125201904&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85125201904&partnerID=8YFLogxK
U2 - 10.1109/BIBM52615.2021.9669644
DO - 10.1109/BIBM52615.2021.9669644
M3 - Conference contribution
AN - SCOPUS:85125201904
T3 - Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021
SP - 2309
EP - 2313
BT - Proceedings - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021
A2 - Huang, Yufei
A2 - Kurgan, Lukasz
A2 - Luo, Feng
A2 - Hu, Xiaohua Tony
A2 - Chen, Yidong
A2 - Dougherty, Edward
A2 - Kloczkowski, Andrzej
A2 - Li, Yaohang
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2021 IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021
Y2 - 9 December 2021 through 12 December 2021
ER -