a VALIANT Effort | July 2024

Dr. Yihao Liu, our first new VALIANT Fellow at Vanderbilt, has established himself as a leading expert in the development of clinically relevant image analysis methodologies. His research focuses on creating advanced tools and algorithms to enable large-scale analysis and facilitate personalized treatment strategies in medical imaging. Dr. Liu’s work at Johns Hopkins University, where he earned his Ph.D. in Electrical and Computer Engineering and advanced to a research scientist role, involved pioneering contributions to optical coherence tomography, representational learning, and magnetic resonance imaging. His projects have emphasized the importance of building reliable and robust systems for healthcare applications, exploring the boundaries of advanced image analysis. ​

“Exploration is something rewarding,” Yihao remarked. “Changing an environment can be a rewarding process, and that’s why I decided to leave Hopkins and join Vanderbilt.” The allure of a new environment and the potential for collaborative opportunities with the medical and engineering schools at Vanderbilt were pivotal in his decision.

Yihao’s vision for the future of VALIANT emphasizes the importance of collaboration and the practical application of advanced tools. “Combining these tools and applying them to the right applications is going to have the most impactful results in the coming years,” he predicted. He also highlighted the advantage of working closely with clinicians to ensure that their research has practical value and real-world impact.

“I talked to several doctors during my interview here,” he noted. The proximity and collaborative spirit between the medical and engineering schools at Vanderbilt have opened up numerous opportunities for impactful projects. “Sometimes during meetings, it feels like we are speaking different languages,” he observed. Overcoming these communication barriers is crucial for successful partnerships between researchers and clinicians.

Yihao encourages the VALIANT community to leverage its unique strengths and focus on impactful, collaborative projects. “We should really take advantage of our collaboration with doctors to ensure our work has practical value,” he urged.

After getting my PhD in Electrical Engineering from the MASI lab at Vanderbilt University in 2016, I started my industrial career at Siemens Healthineers. There, I specialized in developing and translating AI/ML algorithms into commercial solutions, such as auto-contouring for radiation therapy planning and disease detection for computer-aided analysis. Recently, I joined Johnson & Johnson and shifted my focus to drug discovery to speed up the identification and development of new therapeutics through various modalities of biological data.

AI techniques, like ChatGPT, SAM (code) , and DINOv2 (code), have gained lots of popularity with their impressive capabilities in natural language processing and computer vision tasks. These advancements have captivated not only AI practitioners, but also the general public, who are fascinated by the “emergent” behavior of these models, where AI seems to autonomously tackle a wide range of tasks. While much of the buzz focuses on the sophisticated design of these models, it’s crucial to recognize the equally important role of data strategy. ChatGPT, for example, stands out from its competitors largely due to the introduction of Reinforcement Learning from Human Feedback (RLHF). This approach involves collecting a relatively small but specific dataset to align the AI with human preferences, significantly enhancing its performance. Similarly, SAM employs a human-in-the-loop strategy to efficiently gather high-quality data. By collecting 1 billion masks on 11 million images, SAM covers a wide range of common scenarios in computer vision segmentation, ensuring robust performance. DINOv2, on the other hand, curated a dataset of 142 million images from an initial pool of 1.2 billion raw images. This dedicated selection process prioritizes quality over quantity, proving that a more refined, diverse dataset can significantly boost performance compared to simply amassing vast amounts of data.

The AI community today has a spirit of future nostalgia. While it’s clear that advanced AI models are essential, AI practitioners cannot overlook the data quality, diversity, and relevance. Returning to first principles, data guides problem-solving. It is important to effectively curate (code), diversify (code), and generalize (code) the data, ensuring the AI innovation remains grounded in robust data-centric strategies. The VALIANT folks are sitting in a great position with their strengths: great platform, exceptional AI talents, and strong connections with collaborators who generate valuable data. By understanding and properly handling data, they can identify breakthrough points in technical challenges, achieving better performance and efficiency. Recognizing the importance of data will allow them to develop an intentional strategy to collect relevant and diverse data, enabling them to address more broad and generalized problems effectively, and eventually drive high impact advancements in the field.

“What I value most from my time at Vanderbilt is the understanding that everyone is the architect of their own business, from experimental design to career development. Adopting a mindset focused on robust data strategies forms the foundation for building skyscrapers in the booming AI era.”

Our scientists are breaking new ground.

Join us for an immersive week-long AI Summer School from August 12-15, 2024, at Vanderbilt’s Stevenson Hall, focusing on deep learning and AI advancements. Register by August 5th to secure your spot and dive into cutting-edge AI education, hands-on workshops, and expert-led sessions. For program details and registration, visit here.

VALIANT proudly launches its Industry Affiliates Program, fostering dynamic collaborations with industry leaders to support cutting-edge AI research, education, and innovation. Join to gain access to the latest research, exclusive networking opportunities, and impactful collaboration. For details on membership tiers and benefits, visit our program website.