Options
運用深度學習於膝關節損傷核磁共振影像之人工智慧偵測與診斷模型
Other Title
AI-BASED DETECTION AND DIAGNOSIS MODEL FOR KNEE INJURIES USING DEEP LEARNING ON MULTI-SCAN MRI
Type
thesis
Date Issued
2022-06-10
Author(s)
TRUONG NGUYEN KHANH HUNG
Advisor
康峻宏
Subjects
系所名稱：國際醫學研究博士學位學程
Publisher
國際醫學研究博士學位學程
Description
口試委員：郭柏齡 Po-Ling Kuo;王禎麒 Chen-Chie Wang;黎阮國慶 Nguyen Quoc Khanh Le;康峻宏 Jiunn-Horng Kang ;徐繼聖 Gee-Sern Jison Hsu
Abstract
Introduction: Efficient and accurate detection is vital for the diagnosis and treatment of knee injuries. In recent years, there is an increase in interest in deep learning (DL) approaches to detecting knee injuries in magnetic resonance imaging (MRI). Studies have shown that DL models are capable of reaching the same level as human radiologists when it comes to sensitivity and specificity, while at the same time requiring significantly less training time. Current Artificial Intelligent (AI) - based systems are, however, still limited by many different factors, such as unbalanced classes in training data, or the nature of these systems which makes false positives and false negatives almost an inevitability. There are multiple routes for improving upon the existing DL knee injury detection models. As they continue to become more and more advanced, it is expected that the use of these systems will become more popular in the future.
Method: In this study, we create multi models based on machine learning (ML) and DL algorithms to perform classification, recognition, and segmentation tasks on knee MRI. In which the two most important components in the knee joint in this study are the anterior cruciate ligament (ACL) and meniscus.
The first model, based on the DenseNet 121 neural network structure, was used to classify images with or without ACL injury. The dataset includes 799 knee MRI reports from Cho Ray Hospital (Vietnam). These MRI data were obtained from previous work in the hospital, containing knee MRI reports from 5 years (January 1st, 2015 – December 31st, 2019)
Using the Faster-region convolutional neural network (Faster - RCNN) and several convolutional neural networks (CNN) backbone tests, such as VGG-16, Res-Net50, DenseNet-121, EfficientNet-B0, and EfficientNetV2 - B0 algorithms, the second group of models can recognize the ACL on knee MRI as a function of the typical imaging characteristics. This research collected 256 knee MRI examinations performed at Cho Ray Hospital, Ho Chi Minh City, Vietnam, between January 1, 2018, and December 31, 2020 (including training and testing datasets).
The third model focuses on automatic identification and classification of meniscus based on the Yolo-v4 object detection model. At the same time, the lesion location is also shown on images by the GRAD-CAM technique. The total number of subjects used in this study was 704 patients, including meniscus lesions and the control group. All MRIs in this study were collected before the surgery, and all had no prior surgical history. The MRI scanner at Cho Ray Hospital is MAGNETOM Skyra 3T (Siemen), and at Hoan My Hospital is 3.0T MRI Scanners SIGNA (GE Healthcare). In addition, we also used a public dataset - MRNet dataset (validation dataset) from Stanford University Medical Center with 120 examinations for external testing.
Results: The area under the ROC curve (AUC) for the ACL injury classification system was 80.63% with the axial plane and around 78% with both the sagittal and coronal planes, respectively. All sensitivity and specificity point estimates of the proposed ACL injury detection system were all over 96%, indicating no significant differences in diagnostic performance between different planes.
Our DL model detected meniscus tears with 91.4% accuracy on the internal testing dataset, 89.2% accuracy on the external validation dataset, and 79.9% accuracy on the MRNet dataset, respectively. The meniscus tears were visualized by auto-enlarging the detection area and Grad-CAM images.
Conclusion: This report describes the various approaches in knee MRI experiments to provide different AI models for the prediction of knee injuries. The CNN model applied to classify injured ACL images had high sensitivity and specificity, showing that using a simple structured 2D-CNN is more effective for small datasets and can assist non-experts in assessing the assessment of ACL injuries. The proposed model applied to detect meniscus lesions had high accuracy and specificity, showing that our model can assist non-experts in assessing the assessment of meniscus injuries.
Method: In this study, we create multi models based on machine learning (ML) and DL algorithms to perform classification, recognition, and segmentation tasks on knee MRI. In which the two most important components in the knee joint in this study are the anterior cruciate ligament (ACL) and meniscus.
The first model, based on the DenseNet 121 neural network structure, was used to classify images with or without ACL injury. The dataset includes 799 knee MRI reports from Cho Ray Hospital (Vietnam). These MRI data were obtained from previous work in the hospital, containing knee MRI reports from 5 years (January 1st, 2015 – December 31st, 2019)
Using the Faster-region convolutional neural network (Faster - RCNN) and several convolutional neural networks (CNN) backbone tests, such as VGG-16, Res-Net50, DenseNet-121, EfficientNet-B0, and EfficientNetV2 - B0 algorithms, the second group of models can recognize the ACL on knee MRI as a function of the typical imaging characteristics. This research collected 256 knee MRI examinations performed at Cho Ray Hospital, Ho Chi Minh City, Vietnam, between January 1, 2018, and December 31, 2020 (including training and testing datasets).
The third model focuses on automatic identification and classification of meniscus based on the Yolo-v4 object detection model. At the same time, the lesion location is also shown on images by the GRAD-CAM technique. The total number of subjects used in this study was 704 patients, including meniscus lesions and the control group. All MRIs in this study were collected before the surgery, and all had no prior surgical history. The MRI scanner at Cho Ray Hospital is MAGNETOM Skyra 3T (Siemen), and at Hoan My Hospital is 3.0T MRI Scanners SIGNA (GE Healthcare). In addition, we also used a public dataset - MRNet dataset (validation dataset) from Stanford University Medical Center with 120 examinations for external testing.
Results: The area under the ROC curve (AUC) for the ACL injury classification system was 80.63% with the axial plane and around 78% with both the sagittal and coronal planes, respectively. All sensitivity and specificity point estimates of the proposed ACL injury detection system were all over 96%, indicating no significant differences in diagnostic performance between different planes.
Our DL model detected meniscus tears with 91.4% accuracy on the internal testing dataset, 89.2% accuracy on the external validation dataset, and 79.9% accuracy on the MRNet dataset, respectively. The meniscus tears were visualized by auto-enlarging the detection area and Grad-CAM images.
Conclusion: This report describes the various approaches in knee MRI experiments to provide different AI models for the prediction of knee injuries. The CNN model applied to classify injured ACL images had high sensitivity and specificity, showing that using a simple structured 2D-CNN is more effective for small datasets and can assist non-experts in assessing the assessment of ACL injuries. The proposed model applied to detect meniscus lesions had high accuracy and specificity, showing that our model can assist non-experts in assessing the assessment of meniscus injuries.