Introduction:
Ruojun Zhang is applying artificial intelligence (AI) to healthcare, specifically for patients with Hodgkin’s lymphoma. The research focuses on predicting lesion-level treatment responses, helping doctors tailor personalised treatment plans to reduce unnecessary interventions and minimising harmful side effects. This work is conducted in collaboration with the industrial partner, Clatterbridge Cancer Centre.
The challenge
Hodgkin’s lymphoma is a blood cancer that affects the lymphatic system, primarily occurring in young adults aged 20 to 40. The disease manifests as multiple tumour lesions that can appear in various parts of the body, such as the neck, armpit, or groin. Currently, PET/CT imaging is the most common diagnostic tool used to assess the stage of the disease, as it provides both metabolic and anatomical information in 3 dimensions.
Despite the availability of multiple treatment options—chemotherapy, radiotherapy, targeted therapies, and transplants—these interventions come with varying levels of side effects, with damage to bone marrow being one of the most severe. Experts at Clatterbridge Cancer Centre have observed that different lesions within the same patient respond differently to treatments. However, there is currently no method to predict the best treatment for each individual lesion, meaning treatment plans are often based on overall staging rather than lesion-specific responses.
What I am doing
My bachelor's degree in mathematics and MSc in data science provide me with the technical expertise required to address this challenge. My research plan is to develop an AI model with two modules. The first module segments the lesions trained using clinician-provided segmentation data. The second module predicts treatment response by analysing changes before and after treatment, using the outputs of the segmentation module.
In the first year of my PhD, I used the nnU-Net as the segmentation model to detect each tumour lesion from a publicly available PET/CT dataset (source). The dataset includes 145 lymphoma patients. The nnU-net model is a supervised U-net based convolutional neural network framework for segmenting 2 dimensional or 3 dimensional images with single or multiple input modalities. The mean DICE performance of 5 folds cross validation is 0.750. The coefficient of determination of the total metabolic tumour volume is 0.809. After image resampling, I validated the performance on a private Hodgkin’s lymphoma dataset from my industrial partner. By applying mathematical methods such as connected component analysis and bounding box calculation, I can visualise individual tumour lesions to support doctors in making better treatment decisions. Through the rest of my PhD, I will explore medical image registration and prediction techniques for the development of the second module.
The impact of the project:
By introducing AI-based lesion-level treatment response prediction, my research aims to revolutionise treatment planning for Hodgkin’s lymphoma. Instead of relying solely on disease staging, doctors will be able to tailor treatments based on each lesion’s characteristics, reducing unnecessary dosages, treatment cycles, and side effects. Hodgkin’s lymphoma primarily affects young adults aged 20 to 40, a crucial period for career development and family life. With the benefits of targeted treatment, patients will experience better quality of life, improved physical well-being, and greater productivity. This will enable them to return to work and daily activities more quickly, contributing positively to society. Beyond Hodgkin’s lymphoma, this research has broader implications for other cancers and diseases where personalised treatment is critical.
For my industrial partner Clatterbridge Cancer Centre, this research represents a step forward in integrating AI into clinical workflows. The ability to predict treatment responses at the lesion level will set a new standard for precision medicine, positioning them at the forefront of AI-driven oncology advancements. Additionally, by optimising treatment decisions, this approach will significantly reduce costs associated with ineffective treatments and improve resource allocation, ensuring that healthcare services are used more efficiently and equitably.
Working with an industrial partner has enabled me to bridge the gap between AI research and real-world clinical applications. By combining AI innovation with pressing medical challenges, my research not only advances scientific knowledge but also paves the way for more effective, patient-centred cancer treatments.
Keywords: Hodgkin’s lymphoma, medical image segmentation, medical image registration, treatment response prediction
Back to: Centre for Doctoral Training in Distributed Algorithms