Interest in integrating machine learning (ML) models into the healthcare domain has grown rapidly in recent years. Researchers have been designing clinical applications, particularly for medical image interpretation and analysis using Deep Neural Networks (DNNs) and have reported significantly improved accuracies with improved diagnosis speed. However, despite these promising results, many clinicians are still reluctant to fully adopt ML-based diagnostic systems and perform interventions or provide/personalize the treatment based on the computer-aided diagnostic (CAD) system’s recommendation. One major hurdle in the adaptation of these models in clinical settings is the limited availability of clinical data, including medical images (such as computer tomography angiography), electronic health records (radiology finding reports), laboratory test records (blood test results in tabular form), and patient demography (medical and family history of diseases in textual format) for training and testing generalizability of the developed ML models. Other contributing factors such as data privacy concerns, the time and resource-intensive manual annotation, the scarcity of disease-annotated data, and the variations introduced by different imaging modalities obtained from the same manufacturer’s scanners or variations in the same image modalities from different scanner manufacturers, further complicate the adoption of ML system in a clinical setting. A single-trained model cannot adapt to all these variations and other factors. Furthermore, different hospitals use templates and wording for textual reports, and record systems output tabular data in differing formats.  

Researchers have proposed various solutions to try to overcome these challenges. For example, researchers have been using data augmentation techniques to overcome the limited availability of medical image datasets. While these techniques increase the training data for an ML algorithm, they often generate highly correlated images and potentially lead to a non-generalizable trained model. Image resizing is performed to transform the input image so that it can match the image size required by the trained model. This resizing affects (shrinks or enlarges) the anatomy of the region of interest which creates problems of accurate training, diagnosis, and interventions for specific diseases by clinicians. Generative models have been proposed as a solution to overcome data augmentation and data privacy issues. These generative models help in data anonymization, however, most of the studies only generate medical images and do not also consider the generation of corresponding masks or ground truth, which are essential prerequisites for model training. Moreover, efficient training of generative models (e.g.  Generative Adversarial Networks or diffusion models) and generation of synthetic medical data through these models is also complex and time-consuming. 

These multifaceted natured challenges emphasize on the need for collaborative efforts across disciplines, to refine existing solutions and develop novel approaches that can effectively mitigate the challenges incurred during the analysis of medical images.  

 The main objective of this workshop is to explore and advance ML techniques for medical data synthesis and evaluation, focusing specifically on generative models. The workshop aims to deepen our knowledge of: 

(1) generative network implementations for synthetic data generation along with corresponding annotations; and 

(2) customizing generative architectures to make them adaptable to diverse modalities (multimodal models). 

Thus, we aim to bring experts, researchers, and clinicians to discuss and address the critical issues around the adoption of generative models in healthcare for medical image analysis and CAD as well as discuss the knowledge discovery aspect to reduce the time complexities of the model. We invite submissions of original research articles, case studies and innovative solutions to overcome the above-mentioned challenges and advance our knowledge in the field.  

The relevant topics to this include but are not limited to: 

1. Implementation of generative models for medical data synthesis. 

2. Customization of generative architectures for adaptive image synthesis. 

3. Applications of synthetic data generation in medical/clinical data analytics. 

4. Generation of synthetic medical data using 3D models.  

5. Challenges and future directions of generative models in medical domain. 

6. Foundational Models for generating high-quality synthetic data in medical domains with a focus on different modalities (images, tabular and time series datasets). 

7. Challenges and limitations of synthetic data in terms of privacy risks, quality assessment, and generalization to diverse datasets.   

8. Applications of synthetic data in training machine learning models, reducing bias, enhancing privacy, and improving data diversity.  

Contact [talha.iqbal; michael.madden; ihsan.ullah] @universityofgalway.ie  or @insight-centre.org to get more information about the workshop