Tianze Xia

Global Challenge: Data Science/AI/Robotics and Health

Abstract:

The rise of artificial intelligence, especially in deep learning for image classification, shows high promise to aid the diagnosis of skin diseases. However, many models have been created to give a prediction even if there is a lot of uncertainty in the case. This creates an issue for machine learning systems that are not a good representation of medical practice, where there is often uncertainty in a diagnosis.

In this paper, we proposed and evaluated an uncertainty-aware machine learning framework for classifying images for dermatology. For an image dataset containing about 1200 clinical images, we train a Vision Transformer to differentiate between positive, negative, and uncertain cases. Instead of just providing binary classification or multi-class classification, this model specifically allows uncertainty to be a valid information output. We evaluate the model's performance on accuracy, precision, recall, F1-score, and confusion matrix, achieving roughly 82% accuracy for our classification problem.

Our results showed that the inclusion of an "uncertain" class to the model classification increased safety by forcing fewer overly confident misclassifications on uncertain cases and instead reporting "uncertain," which could lead to more human evaluation. The goal of the model is not to get the highest possible accuracy but to create a model that is safe and provides helpful outputs for physicians. This type of model is especially valuable for the application of AI in a clinical setting in a triage or decision support tool. By emphasizing safety and not just accuracy in this problem, we hope to set an example for reliable artificial intelligence.

I focused on addressing the global problem of unequal access to healthcare in underserved environments with this project. My initial thought while considering an AI solution for the problems of limited dermatologist access was to maximize accuracy, but this project changed how I viewed AI: providing useful information even in uncertainty.

By removing overconfidence, I tried to reduce risk that a mistake is made, especially by underprivileged communities who rely on a trusted AI for a proper diagnosis. In the future, I will continue to explore the area of engineering decision-making models and how to combine that with medicine to develop a reliable, safe, and fair system for use in healthcare.

Bio:

Tianze Xia graduated from New York University with a degree in Biomolecular Science and a minor in Mathematics. His time in New York was a formative experience, strengthening his intellectual curiosity while fostering resilience and adaptability.

Born in Ürümqi, China, and raised between Ürümqi and Madrid, Tianze developed a global perspective that shaped his interest in the molecular foundations of human biology. This cross-cultural background continues to inform his analytical approach and academic pursuits.

At NYU, Tianze served as a peer tutor for biology and chemistry at the NYU Polytechnic Tutoring Center, supporting students in challenging subjects and reinforcing his commitment to collaborative learning. He also volunteered at NewYork-Presbyterian Hospital, contributing to patient support while gaining exposure to clinical environments and healthcare systems.

Outside of his academic and professional work, Tianze pursued a range of interests, including weightlifting, playing the piano at 5 MetroTech Center, creating ceramics at Steinhardt’s studio, playing basketball, biking across the Manhattan Bridge, and traveling.