SAVE AS PDF
NORTH CAROLINA OFFICIAL ABSTRACT and CERTIFICATION
Title: Leveraging LLMs and Computer Vision for Skin Lesion Analysis: A Step Select Division,
Toward Explainable AI in Melanoma Risk Assessment Category, Region:
Student Srija Roy Elementary
Name(s): Biological Sci
Chemistry.
School Name: The Early College at Guilford, Greensboro, NC Earth/Env.
City, State: Engineering/
Technology
Abstract (start typing below at the left margin, 250 word maximum length): Physics/Math
Melanomas are the most lethal form of skin cancer and have the potential to become life-
threatening within 6 weeks if left untreated. Survival rate of melanoma drops by 32% if Junior
treatment is delayed beyond stage one. While machine learning techniques have been applied for Senior
early-stage melanoma detection, research on creating explainable models and affordable tools
remain limited. The current project aims to address this by creating a user-friendly web-based Bio Science A
application that utilizes state-of-the-art Large Language Models (LLM) to analyze progression of
a skin lesion over time by using images of the lesion and provide probability estimates of the Bio Science B
skin lesion to be a melanoma. The system uses computer vision (CV) techniques including Chemistry
contour detection, color space analysis, and asymmetry calculations through image bisection on Earth/Env
images obtained from the labeled "Visuals Online" collection of the National Cancer Institute (NIC).
The statistical analysis is based on the ABCD rule (Asymmetry, Border irregularity, Physics
Color variation, and Diameter changes) that is widely used in skin lesion analysis. Statistical Engineering
significance of the observed changes on pairs of images is evaluated using one-sample t-tests
Technology x
against a baseline of no change. For melanoma prediction, the paper uses a variety of open-
sourced LLM-s obtained from the Ollama open-source tool. The web application has been built Math/Data _____
using Streamlit for the frontend interface. The application represents a novel approach to
melanoma risk assessment by combining image analysis techniques with modern LLM Select Region:
capabilities, potentially serving as a cost-effective tool for preliminary screening and monitoring Region 1
of suspicious skin lesions.
Region 2
Region 3A
Region 3B
Region 4
Region 5 x
Region 6
Region 7
Region 8
Please answer the following: Region Vir.
1. As a part of this research project, the student(s) directly handled, manipulated, or interacted
with (check ALL that apply): _____ human participants _____ vertebrate animals _____ potentially
hazardous biological agents _____ microorganisms _____ rDNA _____ tissue _____ x NONE
2. I/we worked or used equipment in a regulated research institution or industrial setting: ______ Yes _______
x No
3. This project is a continuation of previous research. _______ Yes _______
x No
4. My display board includes non-published photographs/visual depictions of
humans (other than myself): _______ Yes _______
x No
Seal provided
5. This abstract describes only procedures performed by me/us, reflects my/our
at State level
own independent research, and represents one year’s work only
of NCSEF
_______
x Yes _______ No.
6. I/we hereby certify that the abstract and responses to the above statements are
correct and properly reflect my/our own work. _______
x Yes _______ No
Revised 8/2023
