Medical Image Understanding and Analysis: 25th Annual Conference, MIUA 2021, Oxford, United Kingdom, July 12–14, 2021, Proceedings

Preface
Organization
Contents
Biomarker Detection
Exploring the Correlation Between Deep Learned and Clinical Features in Melanoma Detection
1 Introduction
2 Dataset and Methodology
2.1 Dataset: Description and Pre-processing
2.2 Deep Architectures
2.3 ABCD Clinical Features and Classification
3 Experiments and Results
3.1 Quantitative Results
3.2 Alignment Between ABCD Features and Deep Learned Features
3.3 Qualitative Results
4 Conclusion
References
An Efficient One-Stage Detector for Real-Time Surgical Tools Detection in Robot-Assisted Surgery
1 Introduction
2 Methodology
2.1 Network Architecture
2.2 Loss Function for Learning
3 Experiment and Results
3.1 Dataset
3.2 Experiment Settings
3.3 Results
4 Conclusion
References
A Comparison of Computer Vision Methods for the Combined Detection of Glaucoma, Diabetic Retinopathy and Cataracts
1 Introduction
2 Research Problem and Environment
3 Literature Review
4 Experiment Setup
5 Model
5.1 Image Acquisition
5.2 The Composition of the Trained Models and Classifiers
5.3 Pre-processing
5.4 Feature Extraction and Classification
5.5 Displaying the Results
6 Results
7 Result Analysis
8 Conclusion
References
Prostate Cancer Detection Using Image-Based Features in Dynamic Contrast Enhanced MRI
1 Introduction
2 Materials and Methods
2.1 Dataset
2.2 Preprocessing
2.3 Feature Extraction
2.4 Classification
3 Results
3.1 Comparison of Image-Based, Pharmacokinetic and Perfusion-Related Features
3.2 Aggregation of All Features
4 Discussion
5 Conclusions
References
Controlling False Positive/Negative Rates for Deep-Learning-Based Prostate Cancer Detection on Multiparametric MR Images
1 Introduction
2 Methods
2.1 Problem Definition
2.2 Overall Training Loss Function
2.3 Lesion-Level Cost-Sensitive Classification Loss
2.4 Slice-Level Cost-Sensitive Classification Loss
3 Experiments and Evaluation
3.1 Data Set and Implementation Details
3.2 Evaluation Metrics
4 Results
4.1 Adjusting Mis-classification Cost at Lesion-Level
4.2 Adjusting Mis-classification Cost at Slice-Level
4.3 Adjusting Mis-classification Cost at Both Levels
4.4 Results Analysis
5 Conclusions
References
Optimising Knee Injury Detection with Spatial Attention and Validating Localisation Ability
1 Introduction
2 Related Work
3 Materials
4 Method
4.1 Model Backbone
4.2 Spatial Attention
4.3 Single-Plane and Multi-plane Analysis
4.4 Training Pipeline
5 Evaluation
5.1 Quantitative
5.2 Ablation Study
6 Explainability
6.1 Localisation Ability
6.2 Features
6.3 Limitations
7 Conclusion
References
Improved Artifact Detection in Endoscopy Imaging Through Profile Pruning
1 Introduction
2 Proposed Method
2.1 Artifact Detection
2.2 Novel Pruning Method Using Instance Profiles
3 Results
3.1 Dataset
3.2 Evaluation Metrics
3.3 Experimental Setup
3.4 Quantitative Results
3.5 Qualitative Results
4 Discussion and Conclusion
References
Automatic Detection of Extra-Cardiac Findings in Cardiovascular Magnetic Resonance
1 Introduction
2 Materials
3 Methods
3.1 Data Pre-processing
3.2 Binary ECF Classification
3.3 Multi-label ECF Classification
3.4 Training
3.5 Statistics
4 Results
4.1 Binary ECF Classification
4.2 Multi-label ECF Classification
5 Discussion and Conclusion
References
Brain-Connectivity Analysis to Differentiate Phasmophobic and Non-phasmophobic: An EEG Study
1 Introduction
2 Principles and Methodologies
2.1 Classical CCM
2.2 Estimating the Direction of Causation Using Conditional Entropy
2.3 Classification Using Kernelized Support Vector Machine
3 Experiments and Results
3.1 Experimental Setup
3.2 Data Preprocessing
3.3 Active Brain Region Selection Usings LORETA
3.4 Effective Connectivity Estimation by CCM Algorithm
3.5 Statistical Analysis Using One-Way ANOVA Test
3.6 Relative Performance Analysis of the Proposed CCM
4 Conclusion
References
Image Registration, and Reconstruction
Virtual Imaging for Patient Information on Radiotherapy Planning and Delivery for Prostate Cancer
1 Introduction
2 Materials and Methods
2.1 Study Design
2.2 Eligibility and Exclusion Criteria
2.3 Radiotherapy
2.4 Bladder and Rectal Measurements
2.5 Bladder Volume Model
2.6 Statistical Analysis
3 Results
4 Discussion
5 Conclusion
References
Data-Driven Speed-of-Sound Reconstruction for Medical Ultrasound: Impacts of Training Data Format and Imperfections on Convergence
1 Introduction
2 Methods
2.1 Data Simulation
2.2 Network Setup
3 Results and Discussion
3.1 Data Format
3.2 Decimation
4 Conclusion
References
Selective Motion Artefact Reduction via Radiomics and k-space Reconstruction for Improving Perivascular Space Quantification in Brain Magnetic Resonance Imaging
1 Introduction
2 Materials and Methods
2.1 Subjects, Magnetic Resonance Imaging and Clinical Visual Scores
2.2 Image Quality Assessment
2.3 Motion Artefact Reduction
2.4 PVS Segmentation
2.5 Comparison Against a Relevant Framework
2.6 Validation Against Clinical Parameters
3 Results
3.1 Image Quality Classification Results
3.2 Motion Artefact Reduction
3.3 Relationship Between Computational Measures and Clinical Visual Scores
4 Discussion
References
Mass Univariate Regression Analysis for Three-Dimensional Liver Image-Derived Phenotypes
1 Introduction
2 Materials and Methods
2.1 Data
2.2 Image Analysis and Mesh Construction
2.3 Mass Univariate Regression Analysis
3 Results
4 Discussion and Conclusions
References
Automatic Re-orientation of 3D Echocardiographic Images in Virtual Reality Using Deep Learning
1 Introduction
1.1 Related Work
2 Methodology
2.1 Data and Labelling
2.2 Methods
3 Results
4 Applications
4.1 Scene Setup
4.2 Integration
5 Discussion
References
A Simulation Study to Estimate Optimum LOR Angular Acceptance for the Image Reconstruction with the Total-Body J-PET
1 Introduction
2 Methods
3 Results
4 Conclusions
References
Optimised Misalignment Correction from Cine MR Slices Using Statistical Shape Model
1 Introduction
2 Preprocessing and Initial Misalignment Corrections
2.1 Preprocessing
2.2 Intensity and Contours Based Misalignment Corrections
3 Proposed Misalignment Correction Using Statistical Shape Model
3.1 Fitting the Statistical Shape Model
3.2 Misalignment Correction Using the SSM
4 Experimental Analysis
5 Conclusion
References
Slice-to-Volume Registration Enables Automated Pancreas MRI Quantification in UK Biobank
1 Introduction
2 Materials and Methods
2.1 UK Biobank Data
2.2 Slice-to-Volume Registration Method
2.3 SVR Implementation and Inference at Scale
2.4 Automated Quality Control
2.5 SVR Validation
3 Results
3.1 T1 Quantification: No Registration vs SVR-SSC
3.2 SVR Validation
4 Discussion and Conclusions
References
Image Segmentation
Deep Learning-Based Landmark Localisation in the Liver for Couinaud Segmentation
1 Introduction
2 Methodology
2.1 Dataset
2.2 Landmark Localisation Model
2.3 Direct Segmentation Model
2.4 Spatial Configuration Post-processing
2.5 Training and Evaluation
3 Results
3.1 Landmarking Accuracy
3.2 Couinaud Segmentation Accuracy
4 Discussion and Conclusion
References
Reproducibility of Retinal Vascular Phenotypes Obtained with Optical Coherence Tomography Angiography: Importance of Vessel Segmentation
1 Introduction
2 Methods
2.1 Participant Demographics and Imaging Protocol
2.2 Image Analysis
2.3 Statistical Analysis
3 Results
3.1 Microvascular Phenotype Reproducibility over Repeated OCTA Imaging
3.2 Dependence of Microvascular Phenotypes on the Choice of Segmentation/Skeletonization Algorithm
4 Discussion
References
Fast Automatic Bone Surface Segmentation in Ultrasound Images Without Machine Learning
1 Introduction
2 Methods
2.1 Simplified Segmentation Method with Bone Probability Map
2.2 Image Acquisition and Hardware Pre-sets
2.3 Algorithm Testing
2.4 Performance Testing Against a Machine Learning Model
3 Results
3.1 Processing Time
3.2 Quantitative Comparison Between Methods
3.3 Qualitative Comparison Between Methods
3.4 Performance Comparison with U-Net
4 Discussion and Conclusion
References
Pancreas Volumetry in UK Biobank: Comparison of Models and Inference at Scale
1 Introduction
2 Materials and Methods
2.1 Data Acquisition
2.2 Data Labelling and Preprocessing
2.3 Model Architectures
2.4 Model Training and Testing
2.5 Model Inference at Scale
3 Results
3.1 Model Evaluation
3.2 Comparison with Volumetry from Pancreas-Specific Scan
3.3 UK Biobank Population Volumetry
3.4 Pancreas Volume Diurnal Variation.
4 Discussion and Conclusion
References
Ensemble of Deep Convolutional Neural Networks with Monte Carlo Dropout Sampling for Automated Image Segmentation Quality Control and Robust Deep Learning Using Small Datasets
1 Introduction
1.1 Related Work
1.2 Contributions
2 Methodology
2.1 Data
2.2 Overview of the Ensemble Framework
2.3 U-Nets with Monte Carlo Dropout
2.4 Combined Segmentation Models
2.5 Prediction of Segmentation Quality
2.6 Evaluation
3 Experimental Results
3.1 Segmentation Performance
3.2 Regression-Based DSC Prediction Accuracy
3.3 Comparison with Monte Carlo-Based DSC Prediction
4 Conclusion
References
Reducing Textural Bias Improves Robustness of Deep Segmentation Models
1 Introduction
2 Materials and Methods
2.1 Dataset
2.2 Textural Noise Simulation
2.3 CNN Architecture
2.4 Experiments
3 Results and Discussion
4 Conclusion
References
Generative Models, Biomedical Simulation and Modelling
HDR-Like Image Generation to Mitigate Adverse Wound Illumination Using Deep Bi-directional Retinex and Exposure Fusion
1 Introduction
2 Related Work
3 Our Approach: Algorithmic Pipeline
4 Evaluation and Results
4.1 Evaluation Experiments
5 Conclusion and Future Work
References
Deep Learning-Based Bias Transferpg for Overcoming Laboratory Differencespg of Microscopic Images
1 Introduction
2 Related Work
3 Methodology
3.1 Generative Approaches
3.2 Data
4 Results and Discussion
4.1 Results
4.2 Discussion
4.3 Limitations
5 Conclusion and Outlook
References
Dense Depth Estimation from Stereo Endoscopy Videos Using Unsupervised Optical Flow Methods
1 Introduction
2 Methods
2.1 Baseline Unsupervised Optical Flow Loss Functions
2.2 Proposed Method
3 Dataset and Implementation
4 Results
4.1 Evaluation Metrics
4.2 Comparison with State-of-the-Art Depth Reconstruction Methods
4.3 Ablation Study
4.4 Comparison with Top Methods in the SCARED Challenge
5 Conclusion
References
Image Augmentation Using a Task Guided Generative Adversarial Network for Age Estimation on Brain MRI
1 Introduction
2 Methodology
2.1 GAN Based Image Synthesis
2.2 Task-Guided Branch
2.3 Latent Space Interpolation for Image Synthesis
3 Experiments and Results
3.1 Dataset
3.2 Experiments
3.3 Results
4 Discussion and Conclusions
References
First Trimester Gaze Pattern Estimation Using Stochastic Augmentation Policy Search for Single Frame Saliency Prediction
1 Introduction
2 Data Augmentation Strategy
2.1 Data
2.2 Data Preparation
2.3 Encoder-Decoder Network
2.4 Mixed-Example Data Augmentation
2.5 Random Augmentation
2.6 Training
2.7 Saliency Map Prediction
2.8 Evaluation Metrics
3 Experiments and Results
3.1 Quantitative Results
3.2 Representative Examples
4 Discussion
5 Conclusion
References
Classification
Dopamine Transporter SPECT Image Classification for Neurodegenerative Parkinsonism via Diffusion Maps and Machine Learning Classifiers
1 Introduction
2 Datasets of PPMI and Clinical Cohort
2.1 PPMI Dataset
2.2 Clinical Dataset of KCGMH-TW
2.3 Labeling Criterion
3 Methodology
3.1 Training Sample Reduction via Diffusion Maps
3.2 Nystrm's Out-Of-Sample Extension
4 Experiments
4.1 Two Steps Model Ensemble and Classifer Selection
4.2 Classification
5 Diagnosis and Discussion
5.1 Two Model Confusion Matrices
5.2 Visualization
6 Conclusion
References
BRAIN2DEPTH: Lightweight CNN Model for Classification of Cognitive States from EEG Recordings
1 Introduction
2 Related Studies
2.1 Cognitive Relevance
2.2 Deep Learning Networks
3 Data and Methods
3.1 Data and Preprocessing
3.2 Methods
4 Results and Discussion
4.1 Performance
4.2 Parameters and Time
4.3 Training vs Performance Tradeoff
4.4 Visualization
4.5 Ablation and Performance Studies
5 Conclusion
References
D'OraCa: Deep Learning-Based Classification of Oral Lesions with Mouth Landmark Guidance for Early Detection of Oral Cancer
1 Introduction
2 Related Work
2.1 Mouth Landmark Detection
2.2 Oral Lesion Classification
3 Methodology
3.1 Mouth Landmark Detection Module
3.2 Oral Lesion Classification Module
4 Experiments
4.1 Dataset and Metrics
4.2 Mouth Landmark Detection Result
4.3 Oral Lesion Classification Result
5 Conclusion
References
Towards Linking CNN Decisions with Cancer Signs for Breast Lesion Classification from Ultrasound Images
1 Introduction
2 Background and Related Work
2.1 Breast Cancer Signs
2.2 CNN Classification Models
2.3 CNN Decision Understanding and Visualization
3 Materials and Methods
3.1 Data Collection
3.2 Breast Lesion Classification Model
3.3 Calcification Cancer Sign Analysis
4 Experimental Results
4.1 Breast Lesion Classification
4.2 Breast Lesion Calcification Analysis
5 Conclusion and Discussion
References
Improving Generalization of ENAS-Based CNN Models for Breast Lesion Classification from Ultrasound Images
1 Introduction
2 Materials and Methods
2.1 Datasets
2.2 Optimizing CNN Architecture with ENAS Approach
2.3 Reducing ENAS Model Generalization Error
3 Experiments and Results
3.1 Evaluating ENAS17 Model Generalization
3.2 Reducing Generalization Error for ENAS Models
3.3 Comparison with Existing Methods
4 Discussion
5 Conclusion
References
Image Enhancement, Quality Assessment, and Data Privacy
Comparison of Privacy-Preserving Distributed Deep Learning Methods in Healthcare
1 Introduction
1.1 Federated Learning
1.2 Split Learning
1.3 SplitFed
2 Related Work
3 Data and Methods
3.1 Data
3.2 Topology and Neural Network Architectures
3.3 Federated Learning Settings
3.4 Split Learning Settings
3.5 SplitFed Learning Settings
3.6 Evaluation Metrics
4 Results
4.1 Classification Performance
4.2 Elapsed Training Time
4.3 Data Communication
4.4 Computation
5 Conclusion
References
MAFIA-CT: MAchine Learning Tool for Image Quality Assessment in Computed Tomography
1 Introduction
2 Materials and Methods
2.1 Data Set and Phantom
2.2 Platform
2.3 Model Used
2.4 Evaluation
3 Results
3.1 Platform Validation
3.2 Synthetic Data
3.3 Measurements
4 Discussion
4.1 Platform Validation
4.2 Model Performance – Synthetic Data
4.3 Model Performance – Measurements
4.4 Limitations
5 Conclusion
References
Echocardiographic Image Quality Assessment Using Deep Neural Networks
1 Introduction
1.1 Overview
1.2 Related Works
1.3 Main Contributions
2 Materials and Methods
2.1 Definition of Legacy Attributes of 2D Image Quality
2.2 Definition of Domain Attributes of 2D Image Quality
2.3 Ground Truth Definition (Expert’s Manual Score Criteria)
2.4 Data Sources
2.5 Network Architecture
2.6 Training, Batch Selection, Data Augmentation
2.7 Evaluation Metrics
3 Results and Analysis
3.1 Study Limitation and Future Work
4 Conclusion
References
Robust Automatic Montaging of Adaptive Optics Flood Illumination Retinal Images
1 Introduction
2 Materials and Methods
2.1 AO Image Acquisition and Problem Framework
2.2 Phase Correlation
2.3 Scale Invariant Feature Transform (SIFT)
2.4 Phase Correlation Followed by Scale Invariant Feature Transform (PC-SIFT)
2.5 Alignment Accuracy
3 Results and Discussion
4 Conclusion
References
Radiomics, Predictive Models, and Quantitative Imaging
End-to-End Deep Learning Vector Autoregressive Prognostic Models to Predict Disease Progression with Uneven Time Intervals
1 Introduction
2 Related Work
2.1 Autoregression
2.2 Explicit Feature Extraction
2.3 Implicit Feature Extraction
3 Method
3.1 Overall Framework
3.2 CNN
3.3 Time Series
3.4 Interval Scaling
3.5 Classification Layer
4 Experiments
4.1 Dataset
4.2 Computation
4.3 Preprocessing
4.4 Metrics
4.5 Results
4.6 Class Activation Maps
5 Conclusions
References
Radiomics-Led Monitoring of Non-small Cell Lung Cancer Patients During Radiotherapy
1 Introduction
2 Overview of the Combined Texture and Level Set Model
2.1 Parallel Level Sets in Vector-Valued Image Model
3 Tests Conducted on NSCLC Cohort
3.1 Registration of CT and CBCT
4 Results and Discussion
4.1 Assessing the Clinical Performance in the Absence of a Ground Truth
4.2 Accessing Performance on Non-medical Data with a Ground Truth
5 Conclusion
References
Deep Learning Classification of Cardiomegaly Using Combined Imaging and Non-imaging ICU Data
1 Introduction
2 Data and Methods
2.1 Datasets
2.2 Preprocessing
2.3 Combining Imaging and Non-imaging Data
2.4 Models
3 Results
4 Discussion
4.1 Principal Findings
4.2 Strengths and Weaknesses of the Study
References
Author Index