Deep Learning for Medical Image Computing

IIDS67562 Deep Learning for Medical Image Computing 15 Semester 2 Postgraduate Taught Level 7 Jinming Duan Unit coordinator Mobarak Hoque Unit coordinator Division of Informatics, Imaging and Data Sciences 1 FHEQ level (Framework for Higher Education Qualifications) ' Masters/Integrated Masters P4 ' European Credit Transfer & Accumulation System Rating : 7.5 The course unit will use lectures, guided exercises, and practical programming sessions to introduce students to the fundamental concepts and techniques of neural networks and its role in automated learning. The module covers key neural network architectures and demonstrates how to design, train, and evaluate large-scale models, with a focus on understanding their learning dynamics and generalisation performance in practical scenarios. Students will also gain hands-on experience in developing and implementing neural networks for real-world medical imaging applications.  The course unit will use lectures, guided exercises, and practical programming sessions to introduce students to the fundamental concepts and techniques of neural networks and its role in automated learning. The module covers key neural network architectures and demonstrates how to design, train, and evaluate large-scale models, with a focus on understanding their learning dynamics and generalisation performance in practical scenarios. Students will also gain hands-on experience in developing and implementing neural networks for real-world medical imaging applications.  This module aims to introduce students to the fundamental principles and techniques of neural networks, emphasising their relevance to a wide range of medical imaging tasks. Students will explore common neural network architectures and examine how they are applied in areas such as image classification, segmentation, detection, registration and reconstruction. The module also places neural networks within the broader context of state-of-the-art machine learning and artificial intelligence methods used in medical imaging research and clinical practice, providing students with both conceptual understanding and practical insight into modern automated learning techniques.  Students should be able to:<ul><li>Analyse the principles of deep learning and neural networks and apply them to medical image computing tasks.<ul><li>Assessed by the formative and summative-1 or –2 tasks</li></ul></li><li>Appraise alternative neural network architectures (e.g., CNNs, transformers, U-Nets, GANs/VAEs/diffusion) for different imaging tasks<ul><li>Assessed by the formative task and summative-1 or –2 tasks</li></ul></li><li>Critically evaluate training and evaluation strategies for neural networks and interpret results to judge whether performance is reliable and meaningful.  <ul><li>Assessed by the formative task</li></ul></li><li>Appraise how data preparation and data quality affect model behaviour and performance and propose improvements to data handling to strengthen robustness and generalisation.<ul><li>Assessed by the formative task </li></ul></li></ul> Students should be able to:<ul><li>Analyse and critically evaluate different neural network architectures and approaches for different tasks.  <ul><li>Assessed by the formative and summative-1 or –2 tasks</li></ul></li><li>Apply problem-solving skills to design and adapt deep learning models for real-world challenges.<ul><li>Assessed by the formative task</li></ul></li><li>Develop the ability to critically appraise research literature and emerging methods in machine learning and neural networks.<ul><li>Assessed by the formative and summative-1 or –2 tasks </li></ul></li></ul> Students should be able to: <ul><li>Solve moderately complex problems.<ul><li>Assessed by the formative task</li></ul></li><li>Apply neural computation algorithms to specific technical and scientific problems<ul><li>Assessed by the formative task</li></ul></li><li>Use programming language (Python) for numerical problem solving and run advanced program scripts<ul><li>Assessed by the formative task </li></ul></li></ul> Students should be able to:<ul><li>Critically reflect on and strengthen confidence in tackling data-driven and computational problem-solving tasks.<ul><li>Assessed by the formative task, summative-1 and  –2 tasks </li></ul></li></ul> The contents we cover in this unit are:<ol><li>Convolutional Neural Networks (CNNs)<ol><li>Convolutional layers, feature maps, output size, number of parameters, transpose convolution; applications: medical image classification, localisation, object detection, semantic segmentation; data preprocessing.</li></ol></li><li>Longitudinal and Sequential Modeling: <ol><li>RNN fundamentals, LSTM mechanics, Mamba for long sequences, BPTT and gradient dynamics, memory and attention in sequences, sequence representation learning</li></ol></li><li>Foundation Models: <ol><li>Self-supervision, contrastive learning, large scale pretraining, representation learning, adaptation and transfer</li></ol></li><li>Vision Language Foundation Models: <ol><li>Core VLM architectures, image text alignment, cross modal attention, multimodal pretraining, grounded reasoning</li></ol></li><li>Responsible and Trustworthy AI: <ol><li>Interpretability methods, bias and fairness, explainable decisions, visual analytics, uncertainty in medical imaging</li></ol></li><li>Auto-Encoders and VAEs<ol><li>Standard auto-encoders (AEs) and VAEs, encoding and decoding, training, feature learning, pre-training classifiers, limitations and use cases.</li></ol></li><li>Generative Models<ol><li>Variational auto-encoders (VAEs): architecture, training, data generation, interpolation, manipulation; Generative Adversarial Networks (GANs): Generator, Discriminator, adversarial training, loss modifications, advanced GAN architectures.</li></ol></li><li>Transformers<ol><li>Attention mechanism, auto-regressive generative modeling, non-convolutional, non-recurrent architectures; applications in NLP and broader ML tasks.</li></ol></li><li>Diffusion Models<ol><li>Forward and reverse diffusion, loss functions, model training and sampling, applications to medical image generation, text-to-image, VQAs, super-resolution, latent diffusion models.</li></ol></li></ol>• Introduction to programming:Advanced GPU programming in Python using PyTorch; Automatic differentiation and computational graphs; Training neural networks on GPU; Handling datasets and batch processing; Using Python for data preprocessing, visualisation, and model evaluation  The course unit will run over 11 weeks, with the final week dedicated to a quiz-based assessment. No new material will be introduced in that week; instead, students will complete the test and have time to revise the content covered so far. Each week includes in-person lectures, lab practicals and tutorials, and self-directed study supported by recommended reading materials. Lab practicals and tutorials, delivered using PyTorch, will also provide opportunities for formative assessment.  All learning materials are available on Canvas, with each weekly section starting with an introduction to the topics covered and guidance on the recommended study order. Each topic typically includes lecture slides, video recordings, reading materials, an exercise and a practical lab or tutorial. Materials not marked as “optional” are compulsory and may be assessed in the final exam. Lab practical and tutorials can be accessed via Canvas, while Microsoft Teams is used to address module-related questions; students should enrol in the unit’s Team to participate. The Announcements section on Canvas provides important updates from the unit staff, such as changes to deadlines, and should be checked regularly.  Each week includes a session on Python GPU programming, allowing students to practice exercises related to the lecture topics; these sessions are formative and do not carry summative assessment.  1 Written exam 80% 9 Set exercise 20% Further assesment details below: <figure class="table"><table><tbody><tr><td>Assessment task</td><td>Length</td><td>Weighting within unit</td></tr><tr><td> Formative:Lab practicals (Jupyter Notebook) </td><td>2 hours weekly</td><td>0%</td></tr><tr><td>Summative-1:Continuous assessment in Week 12 (via Canvas quiz: multiple-choice questions). </td><td>1 hour</td><td>20%</td></tr><tr><td>Summative-2:Final exam (closed-book exam).    </td><td>2 hours</td><td>80%</td></tr></tbody></table></figure> Formative: In class/On canvas/Teams. Feedback will be provided within the next week Summative-1: Marks and feedback will be provided on Canvas within the test week Summative-2: Marks back within 3 weeks  MSc Applied AI for Med Imaging MSc Applied AI for Med Imaging Not Set Mandatory N  <ol><li>Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT Press.</li><li>Bishop, C. M., & Bishop, H. (2023). Deep learning: Foundations and concepts. Springer Nature.</li><li>Zhou, S. K., Greenspan, H., & Shen, D. (Eds.). (2023). Deep learning for medical image analysis (2nd ed.). Springer. </li></ol> Lectures 33 Seminars 22 95