Project # 51
Efficient Deep Learning: Fine-Tuning lightweight models for Melanoma detection.

Team Lead:	Thanh Tran
Other Team Members:
Faculty Advisor(s):	Dr. Aera Kim LeBoulluec
Department:	IMSE
Project Type:	REU
Title:	Efficient Deep Learning: Fine-Tuning lightweight models for Melanoma detection.
Abstract:	Artificial Intelligence has significantly advanced melanoma detection in medical imaging. Still, the conventional approach demands computationally intensive deep-learning models and substantial high-quality data for effective training. However, this study introduces an efficient alternative by fine-tuning lightweight, pre-trained CNN models, such as EfficientNetB0, for melanoma detection from dermoscopy images. This approach achieves state-of-the-art performance and significantly reduces computational demands, making it a cost-effective solution. The model's effectiveness in melanoma prediction achieves a best average F1 score of 0.9982 across five-fold, with precision and recall scores of 0.99 on our test set, all within an hour of training. We utilize the 2019 ISIC dataset, which presents significant class imbalances between melanoma and non-melanoma cases, and then we split the data into training and validation. Our method outperforms prior studies in melanoma classification while utilizing lightweight models, thereby showing that AI integration into medical analysis is worth investing in, which can benefit patients and professionals.
Poster:	View Poster #51 (PDF)
Judges Scoring:	Submit Evaluation (judges only)

Project # 51Efficient Deep Learning: Fine-Tuning lightweight models for Melanoma detection.