Student Research

Ethan Hong, Zhixin Liang, Yanling Lin, Zachary Po, Sophia Yuxin Zhang, Hamed Hekmat (Research Mentor), and Roxana Daneshjou (Faculty)

This project evaluated the performance and fairness of several AI models for skin cancer detection across different skin tones using the Stanford Diverse Dermatology Images (DDI) dataset. Students compared multiple pretrained models and training approaches to examine both diagnostic performance and disparities across skin tone groups. Results showed that pretrained models generally outperformed models trained from scratch and highlighted important tradeoffs between malignant lesion detection and fairness across groups. The work underscores the importance of diverse datasets and fairness-aware approaches when developing AI tools for dermatology.

Lexi Dai, Danya Jeroudi, Lina Jeroudi, Johanna Kim, Addison Lee, Charlie Peng, Amari Sims, Alaa Youssef (Research Mentor), Curtis Langlotz (Faculty)

This project examined access to AI education across U.S. high schools and the factors associated with differences in availability. Students analyzed a national sample of 184 public and private high schools, looking at AI courses, clubs, school policies, geography, demographics, and school type. The study found substantial disparities in access, including significantly greater availability of AI programs at private schools than public schools and differences across regions and communities. The findings highlight opportunities to expand equitable access to AI education and better prepare students for an increasingly AI-driven future.

Leo Bouchon, Veda Kumar, Amelie Rhew, Ashraya Siruvole, Sarah Zhou, Cally Lin (Research Mentor), and Curtis Langlotz (Faculty)

This project developed an AI-powered chest X-ray screening system designed to detect 14 conditions simultaneously, including pneumonia, edema, atelectasis, pneumothorax, and pleural effusion. Using the CheXpert Plus dataset of more than 224,000 chest X-rays, students compared two AI models to assess both diagnostic performance and computational efficiency. The results demonstrated strong predictive performance and highlighted different strengths across the models, supporting the potential for AI-assisted screening tools to help radiologists identify disease earlier and reduce missed or delayed diagnoses.

Anika Ganu, Joey Guo, Sarah Li, Dylan Po, Elijah Renner; Rayan Ansari and Alaa Youssef (Research Mentors); Bao Do and Curtis Langlotz (Faculty)

This project developed methods to make AI-generated chest X-ray reports more reliable and trustworthy. Students focused on two key challenges: AI-generated findings that are not supported by the image and confidence estimates that do not accurately reflect model performance. Using more than 25,000 chest X-rays and reports, they developed approaches to better calibrate confidence and verify generated findings against image features. The methods improved report accuracy and reduced unsupported findings, demonstrating a potential approach for making vision-language models safer and more useful in radiology.

Arjun Deshpande, Nirvana Choudhury, Jocelyn Louie, Karina Khilnani, Richard Shan, Cally Lin (Research Mentor), and Bao Do (Faculty)

This project developed an AI model to detect abnormalities in musculoskeletal X-rays, with the goal of supporting faster diagnosis and reducing missed fractures. Using the Stanford MURA dataset of more than 40,000 radiographic images, students combined two different AI architectures—a convolutional neural network and a vision transformer—to take advantage of their complementary strengths. The combined model outperformed the individual models tested, demonstrating the potential of AI-assisted tools to help prioritize abnormal studies and reduce diagnostic delays, particularly in healthcare settings with limited access to radiology expertise.

Srivatsav Kannan, Sara Pradhan, Anagha Vippagunta, Achuth Vinay, Evan Zhao, Hamed Hekmat (Research Mentor), and Magdalini Paschali (Research Advisor)

This project developed a framework to identify and reduce demographic disparities in medical AI models. Students created a fairness auditing system to measure differences in model performance across patient groups and a debiasing approach designed to reduce those differences while preserving diagnostic performance. Tested on dermatology and chest X-ray datasets, the approach meaningfully improved fairness across both applications with only modest changes in overall model performance, demonstrating a practical strategy for developing more equitable medical AI systems.

Emily He, Dean Jordan, Rayna Kumar, Rishabh Rao, Alex Su, Rayan Ansari (Research Mentor), Sergios Gatidis (Faculty)

This project developed an AI system to assess heart function in children using just two frames from an echocardiogram, with the goal of making cardiac assessment more reliable with handheld point-of-care ultrasound. Using nearly 3,000 pediatric echocardiograms, students trained and evaluated a model to estimate ejection fraction and tested its performance under simulated motion and image-quality degradation. The system remained accurate under these challenging conditions, demonstrating the potential for AI to support pediatric heart assessment in bedside and resource-limited settings.