Shuohan Xue (Roy)
🎓 PhD in Electronic and Electrical Engineering, University of Sheffield
📬 Email Me | LinkedIn | IEEE Xplore
🧠 Researcher in Adversarial machine/deep learning, Robust machine/deep learning, medical imaging, and XAI
- Robust and interpretable deep learning for medical and clinical applications
- Adversarial example detection and defence under distribution shifts
- 3D medical image analysis using volumetric CT (COVID-CT-MD) and ROI-aware models
- Quantitative evaluation of model uncertainty and adversarial awareness
- Hybrid feature modelling using deep neural networks and wavelet scattering transforms
- Knowledge transfer and hybrid feature modelling
- Signal and image processing for healthcare and biomedical systems
I recently completed my PhD in Electronic and Electrical Engineering at the University of Sheffield. My research focused on developing robust and interpretable deep learning models for real-world medical image analysis, with an emphasis on detecting adversarial examples, uncertainty-aware decision-making, and achieving feature-level robustness under distribution shifts.
In my thesis, I proposed a set of quantisation, regularisation, and knowledge transfer methods using wavelet scattering features to enhance the adversarial resilience of deep neural networks. I also developed a novel region-of-interest aware 3D ResNet architecture for prognostic classification of volumetric chest CT scans during the COVID-19 pandemic.
I have published four peer-reviewed papers (IEEE ICASSP, IEEE ACCESS, BMVC, and one under review at IEEE TNNLS), with over 70 citations to date. I have experience working with interdisciplinary teams across engineering, clinical medicine, and computer science, and have contributed to project supervision and collaborative research development. During my research, I have worked extensively with datasets such as COVID-CT-MD for volumetric chest CT classification, BraTS for brain tumour segmentation, ImageNet for large-scale visual recognition, Cityscapes for urban scene understanding, and CamVid for semantic segmentation in driving environments.
1. 🫁 Covid-19 Diagnosis with 3D Transfer Learning [PDF]
We developed a 3D deep transfer learning pipeline for classifying COVID-19 infections from volumetric CT scans. Using pretrained 3D ResNet architectures, the system achieved high classification performance with minimal training data, demonstrating potential for real-world clinical triage support during the pandemic.
Tech stack: PyTorch, MONAI, 3D ResNet, Medical image pre-processing
Published in: IEEE ICASSP 2021
2. 🧠 ROI-Aware 3D ResNet for Chest CT Classification [PDF]
This work proposes an ROI-aware feature enhancement block integrated into a 3D ResNet, focusing on regions affected by lung infection. The model outperforms conventional 3D ResNets in COVID-19 CT classification by improving sensitivity to disease-relevant areas.
Tech stack: PyTorch, SimpleITK, lung segmentation, attention-based pooling
Published in: IEEE ACCESS (2023)
This paper validates the inherent robustness of hand-crafted features (especially WSN) under adversarial attacks and proposes:
- Adversarial Awareness Score (AAS): A feature-discrepancy metric to detect adversarial examples.
- AAS-Guided Jacobian Regularisation: A novel training strategy enhancing robustness without sacrificing clean accuracy.
Tech stack: MATLAB, PyTorch, ImageNet experiments, adversarial attacks Published in: BMVC 2025
We present a neural projection method that maps WSN features to DNN feature space, enabling robust classification from adversarially stable representations. The technique enhances robustness by transforming input features before classification, without modifying the DNN architecture.
Tech stack: PyTorch, Autoencoders, Seq2Seq regression, Wavelet Scattering Network
Status: Under review at IEEE Transactions on Neural Networks and Learning Systems (TNNLS)
- "Covid-19 Diagnostic Using 3D Deep Transfer Learning"
IEEE ICASSP 2021 - "ROI-Aware 3D ResNet for COVID-19 CT Scan Classification"
IEEE ACCESS 2023 - "Validating the Adversarial Robustness of Hand-Crafted Features"
BMVC 2025 (Accepted) - "Adversarial Feature Projection Network for Robust Image Classification"
IEEE TNNLS (Under Review)
📊 Citations: 70+ (as of June 2025)
- Deep Learning: PyTorch, TensorFlow, Keras, MONAI
- Classical ML & Signal Processing: MATLAB, Scikit-learn
- Medical Imaging: CT/MRI processing, DICOM
- Adversarial ML: FGSM, PGD, CW attacks, detection & defence methods
- Software Engineering: Git, GitHub Actions, Python OOP, unit testing
- Writing & Collaboration: LaTeX, Overleaf, academic writing, grant drafting
Foxconn Technology Group
Jul 2015 – Jul 2017
- Maintained calibration systems for iPhone/iPad camera production
- Integrated embedded hardware and optical sensors in mass production
- Left role to pursue MSc and PhD in the UK
- 📧 Email: sxue6@sheffield.ac.uk
- 🌐 Personal Site: GitHub Pages
- 🧪 Google Scholar
📍 Always open to collaborations in machine learning, healthcare AI, and neuroscience research.