|
辨識腦腫瘤的類別對於治療計畫的制定具有重要作用。雖然病理診斷是腦瘤類別判定的黃金標準，但它需要花費比較多時間，而利用機器學習由影像進行間接的診斷，則提供了另一條更加快速的方法。此外，臨床資料常具有的類別不平衡特性會影響機器學習的表現。因此，本研究著重於探討不同手段對於卷積神經網路在腦瘤磁振影像進行類別判定的效果，以期在臨床應用上能夠有穩健且精確的表現。
本研究使用的資料來自臺北榮民總醫院，包含六種常見的腦瘤類型（聽神經瘤、神經膠質瘤、腦膜瘤、腦轉移瘤、腦下垂體瘤、神經鞘瘤）以及無腦瘤個案，並採用有打顯影劑的T1權重橫切面磁振影像。實驗方法上，我們利用不同訓練資料，包括1.只有磁振影像，2.前項1再增加腫瘤遮罩，3.前項1再增加腫瘤區域影像等三種，以及不同訓練方式，包括1.不平衡的訓練資料，2.隨機抽樣的平衡訓練資料，3.固定的平衡訓練資料等三種，來訓練改良的卷積神經網路模型，並針對同位個案的二維影像結果進行兩種投票，包含最大投票法和權重投票法，得出最終個案層級的分類結果。
實驗結果顯示，利用增加腫瘤區域影像，且採用隨機抽樣的平衡訓練資料，在權重投票法的組合下有最佳的結果，平均敏感度達到95.34%，除了腦膜瘤的敏感度是76.42%之外，其他類別都達到93%以上。進一步畫出熱圖，也發現多數情況模型學習到的分類依據是合理的。先較於先前研究，本研究提出的腦瘤分類模型更貼近臨床實際情境，能處理類別不平衡情況，並同時區分出多達七種的類別。
The identification of brain tumor types is important for treatment planning. Although pathological diagnosis is considered as the gold standard, it is time consuming. In contrast, and the use of machine learning for image classification provides a faster alternative. However, the imbalanced nature of clinical data can affect the performance of machine learning. Therefore, this study focuses on exploring different training strategies and arrangements of training datasets to classify MR images of brain tumors by convolutional neural networks, with the aim of achieving robust and accurate performance in clinical applications.
The data used in this study were collected from the Taipei Veterans General Hospital and included six common types of brain tumors, namely, vestibular schwannoma, glioma, meningioma, metastasis, pituitary adenoma, and schwannoma, and normal cases (those without brain tumors). In the experiments, we respectively trained the modified-VGG16 model using three different training datasets, which were MRI only, MRI with additional tumor mask, and MRI with additional tumor image, combined with three different strategies, including imbalanced data, randomly-selected balanced data, and fixed balanced data. The classification results using the 2D image slices of the same patients, referred to as the “slice-level” results, were in turn used to derive the “patient-level” classification results via the proposed majority voting method and weighted voting method.
The best patient-level result was via the weighted voting method from the model trained by the randomly-selected balanced training data with additional tumor image, achieving an averaged sensitivity of 95.34%, except for meningioma, where the sensitivity was 76.42%, and over 93% for all the others categories. Furthermore, heatmaps of the best model on testing data showed that the traits learned by the model was reasonable. Compared to previous studies, the proposed brain tumor classification model is able to meet the actual clinical scenario with class-imbalanced data and capable of distinguishing up to seven categories.
中文摘要 i
Abstract iii
Contents v
List of Figures vii
List of Tables ix
Chapter 1. Introduction 1
1.1 Brain Tumors 1
1.2 Related works 2
1.3 The Aim of This Study 4
1.4 Challenges of This Study 5
Chapter 2. Materials and Methods 6
2.1 Dataset Collection 6
2.2 Dataset Preparation 7
2.2.1 Image Preprocessing 7
2.2.2 Dataset Splitting 8
2.2.3 Slice Selection 9
2.3 Deep Learning Model Architecture 12
2.4 Training Stage 13
2.4.1 Hyper-parameters Setting 13
2.4.2 Data Augmentation 14
2.4.3 Experiments of Different Training Techniques 14
2.4.4 Experiments of Providing More Information to Models 16
2.5 Evaluation Stage 17
2.5.1 Evaluation metrics 17
2.5.2 Voting Methods 18
2.6 Software and Hardware Setup 20
Chapter 3. Results 21
3.1 Results of Comparing Different Training Methods 21
3.2 Testing Results of Different Models 25
3.2.1 Testing Results of Models Trained with Different Methods 26
3.2.2 Testing Results of Providing More Information to Models 28
3.3 Testing Results of Adopting Voting Methods 30
3.3.1 Testing Results by Majority Voting 30
3.3.2 Testing Results by Weighted Voting 33
3.3.3 Compare Testing Results of Different Voting Methods 36
Chapter 4. Discussion 40
4.1 Heatmaps of Testing Data for Visualizing What Models Had Learned 40
4.1.1 Necessities of Providing Additional Information of Tumor regions 40
4.1.2 Improvements Made from Voting Methods 46
4.2 Clinical Perspectives for the Classification of Brain Tumors 49
Chapter 5. Conclusion 53
5.1 Conclusion 53
5.2 Future Studies 54
References 56
Appendix A 58
Appendix B 59
Appendix C 60
Appendix D 61
Appendix E 62
1. Brain Tumor: Statistics. [cited 2021 April 29]; Available from: https://www.cancer.net/cancer-types/brain-tumor/statistics.
2. CANCER REGISTRY ANNUAL REPORT, 2018, TAIWAN. Health Promotion Administration, Ministry of Health and Welfare. p. 94-97.
3. Brain Tumors. [cited 2021 Aril 28]; Available from: https://www.aans.org/en/Patients/Neurosurgical-Conditions-and-Treatments/Brain-Tumors.
4. Brain Tumor: Introduction. [cited 2021 April 29]; Available from: https://www.cancer.net/cancer-types/brain-tumor/introduction.
5. 2020 Metastatic Brain Tumors Brochure. AMERICAN BRAIN TUMOR ASSOCIATION.
6. Survival Rates for Selected Adult Brain and Spinal Cord Tumors. 2021 [cited 2021 April 29]; Available from: https://www.cancer.org/cancer/brain-spinal-cord-tumors-adults/detection-diagnosis-staging/survival-rates.html.
7. Cheng, J., Brain tumor dataset. 2017.
8. Swati, Z.N.K., et al., Brain tumor classification for MR images using transfer learning and fine-tuning. Computerized Medical Imaging and Graphics, 2019. 75: p. 34-46.
9. Deepak, S. and P.M. Ameer, Brain tumor classification using deep CNN features via transfer learning. Computers in Biology and Medicine, 2019. 111: p. 103345.
10. Chelghoum, R., et al. Transfer Learning Using Convolutional Neural Network Architectures for Brain Tumor Classification from MRI Images. 2020. Cham: Springer International Publishing.
11. Muhammad Sajjad, S.K., Khan Muhammad, Wanqing Wu, Amin Ullah, Sung Wook Baik, Multi-Grade Brain Tumor Classification using Deep CNN with Extensive Data Augmentation. 2018.
12. DICOM - Overview. [cited 2021 June 15]; Available from: https://www.dicomstandard.org/using.
13. Nyström, L., Inter-hospital brain tumour diagnostics using Private Federated Learning An empirical analysis of convergence in a heterogeneous, non-IID setting and a theoretical review of privacy mechanisms. 2020.
14. Zhang, J., et al., Attention Gate ResU-Net for automatic MRI brain tumor segmentation. 2020. 8: p. 58533-58545.
15. Tang, H., A. Ortis, and S. Battiato. The Impact of Padding on Image Classification by Using Pre-trained Convolutional Neural Networks. 2019. Cham: Springer International Publishing.
16. Lin, M., Q. Chen, and S.J.a.p.a. Yan, Network in network. 2013.
17. Grandini, M., E. Bagli, and G.J.a.p.a. Visani, Metrics for Multi-Class Classification: an Overview. 2020.
18. Ge, C., et al., Enlarged training dataset by pairwise gans for molecular-based brain tumor classification. 2020. 8: p. 22560-22570.
19. Petascale Supercomputer. [cited 2021 May 6]; Available from: https://iservice.nchc.org.tw/nchc_service/nchc_service_hpc.php.
20. Chougule, M., Intra-axial/Extra-axial Brain Tumors, in Neuropathology of Brain Tumors with Radiologic Correlates. 2020, Springer Singapore: Singapore. p. 357-358.
|