I attended the MIDL 2024 conference, held in Paris from July 3rd to 5th, as the sole representative from the TIA Centre. My primary goal was to engage in learning and networking. The event took place at the Sorbonne Université - Pierre et Marie Curie Campus, where I spent a productive week.

Day 1: PhD Symposium

The conference began with the PhD Symposium, providing an opportunity to connect with colleagues from various countries, including Australia, Germany, and France. We participated in a lecture on communication in science and a round table discussion on women and minorities in research. In the afternoon, I enjoyed a treasure hunt activities with the other PhD students around some of Paris's most famous landmarks.

Day 2: Segmentation and Unsupervised Learning

The second day focused on segmentation, a critical topic in my area of study, the nuclei segmentation task in computational pathology. I was particularly interested in presentations on fast nuclei segmentation and an unsupervised novel nuclei segmentation model. It was exciting to engage with scholars worldwide working in similar fields. Additionally, I learned from the broader biomedical image analysis field, discovering transferable ideas and technologies. The keynote session featured Gaël Varoquaux, who discussed integrating machine learning into healthcare. He highlighted the importance of causal thinking in decision-making and the challenges in validating predictive clinical models, including the gaps between analytical and clinical validation.

Day 3: Domain Adaptation and Synthesis

On the third day, I engaged in discussions with poster presenters, exploring new ideas about domain generalization methods for histology image analysis using synthetic histopathology images. I learned valuable quality control methods in computational pathology that can create diverse datasets, useful in the era of foundation models. Dr. Laure Fournier, the keynote speaker, emphasized the challenges in integrating AI into clinical imaging, such as identifying relevant applications, ensuring clinical validation, addressing ethical concerns, and establishing a sustainable economic model.

Day 4: Foundation Models, Explainable AI, and Uncertainty

The final day was dedicated to discussions on foundation models in biomedical image analysis and histopathology. While these models demonstrate impressive capabilities, current evaluations are limited and lack practical guidelines. The keynote speaker, Karim Lekadir, stressed the importance of building trustworthy AI in medical imaging. He argued that beyond technical advancements, key factors like stakeholder engagement, diversity and inclusion, risk management, comprehensive validation, and continuous monitoring are essential for developing robust, safe, and ethical AI tools.

Personal Takeaway

As a second-year PhD student, attending the MIDL 2024 conference was an invaluable experience. Meeting scholars from around the world reassured me that I am not alone in my research direction, which is crucial for learning to collaborate effectively. I am grateful to the MIDL conference for creating a comprehensive agenda, from icebreakers to social events, allowing me to meet new friends and potential collaborators. Last but not least, I deeply appreciate my supervisor, Prof. Nasir Rajpoot, for suggesting that I attend this conference, which inspired me and provided a clearer vision for my future work.

Kesi Xu