Projects

Personalized learning by decoupling global-local mapping

To address the challenge of learning a single global mapping that fits all samples in cross-modality medical image synthesis, this work introduces a sample-adaptive generative adversarial network (GAN) framework. Traditional GANs learn one global mapping from a source modality to a target modality and apply it uniformly, which struggles with the diversity and limited size of medical datasets. The proposed model integrates two complementary learning paths: a global path that captures overall modality translation and a local, sample-adaptive path that focuses on each image’s neighbourhood to extract its distinctive structural and textural characteristics. By jointly considering population-level and subject-specific information, the framework allows the generator to dynamically adjust to individual samples, yielding more accurate and realistic cross-modality synthesis.

Personalize intensity look-up table

To address the challenge of intensity inconsistencies in MRI that hinder accurate brain tumour segmentation, this work proposes a sample-adaptive intensity lookup table (SA-LuT-Net) framework. The method learns a personalised, nonlinear mapping that dynamically adjusts the intensity contrast of each MR image to better align with the segmentation objective. By jointly training a LuT module and a segmentation network, the model enables end-to-end adaptation between individual image characteristics and segmentation performance. By parameterising the intensity mapping function and learning its coefficients per image, the model captures meaningful segmentation-related cues while maintaining interpretability—offering a general strategy for adapting image contrast to improving model robustness across heterogeneous MRI data.

Personalized learning by instance-adaptive supervision

To address the limitations of treating all unlabeled data equally in semi-supervised semantic segmentation, this work proposes an instance-specific and model-adaptive supervision (iMAS) framework. The method recognises that unlabeled samples differ in difficulty and informativeness, and it adapts the learning process accordingly. iMAS evaluates each unlabeled instance’s hardness using a class-weighted intersection-over-union metric and assigns adaptive weights to its consistency loss, allowing the model to focus more effectively on challenging examples as training evolves. In addition, the augmentation strength for each instance is dynamically adjusted based on the model’s generalisation capability, enabling progressive and stable learning. Without introducing extra loss terms or training phases, this instance-adaptive supervision strategy provides an efficient and principled way to align model evolution with data diversity, enhancing the robustness of semi-supervised segmentation.

Personalized learning by instance-adaptive augmentation

To enhance semi-supervised learning, this work introduces a sample-adaptive augmentation (SAA) framework that tailors data augmentation strength based on each sample’s learning state. The method identifies “naive” samples—those already well learned and producing near-zero losses—and augments them more diversely to continue providing meaningful learning signals. SAA consists of two lightweight components: a sample selection module that detects naive samples using historical training information, and a sample augmentation module that applies adaptive, diverse transformations to these samples. By dynamically allocating augmentation intensity according to sample difficulty, SAA maximises the utility of unlabeled data while remaining simple to implement, offering a general strategy for improving data efficiency and model generalisation in semi-supervised learning.

Personalized learning by pixel-wise confidence modeling 

To address the challenges of pixel-dependent noisy labels and class imbalance in medical image segmentation, this work proposes a Collaborative Learning with Curriculum Selection (CLCS) framework. Extending the idea of personalised learning, CLCS adapts supervision at the pixel (sample) levels to match the data quality and model’s evolving capability. The Curriculum Noisy Label Selection (CNS) module dynamically identifies reliable pixels through collaborative voting between two networks, enabling  pixel-specific supervision that evolves over time. Meanwhile, the Noise Balance Loss (NBL) module adjusts the learning weight of uncertain pixels rather than discarding them, ensuring balanced and data-efficient training. This adaptive strategy personalises learning to data characteristics and model maturity, providing a robust approach for handling noisy and imbalanced medical data.

Personalized learning by adaptive codebook-based representation

To bridge the gap between explicit and implicit 3D surface representations, this work introduces Neural Vector Fields (NVF) — a hybrid framework that combines the geometric precision of explicit mesh manipulation with the flexibility of implicit surface learning. NVF directly predicts displacement vectors from surface queries, simultaneously encoding distance and direction fields to enable differentiation-free and topology-agnostic surface reconstruction. Extending the concept of personalised learning to 3D shape representation, NVF employs a soft codebook–based adaptive mechanism, where learned geometric codes are linearly combined with input-specific weights to tailor feature representations for each object. This personalised representation strategy allows NVF to adapt to individual shape characteristics while maintaining generalisation across categories, offering an interpretable and geometry-consistent framework for adaptive 3D reconstruction.