Banca de QUALIFICAÇÃO: ALAN KLINGER SOUSA ALVES

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : ALAN KLINGER SOUSA ALVES
DATE: 31/08/2026
TIME: 09:30
LOCAL: Online
TITLE:

UMergeNet: A New Lightweight U-Net Architecture

KEY WORDS:

Medical image segmentation, Lightweight CNNs, Efficient dense connections. 

PAGES: 102
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SPECIALTY: Processamento Gráfico (Graphics)
SUMMARY:

This work presents UMergeNet, a lightweight segmentation architecture with only 1.7 million trainable parameters. UMergeNet was evaluated across five datasets and compared against eleven state-of-the-art \gls{cnn} architectures. Across all datasets, it achieved a mean F1-score of 94.65\% and a \gls{miou} of 90.13\%, ranking second overall. UMergeNet achieved 1,208 \gls{fps} on a \gls{gpu}, 11.84 \gls{fps} on a \gls{cpu}, and required only 0.88 \gls{gflops}. Compared with U-Net 3+, the top-performing model, it uses 93.60\% fewer trainable parameters and 98.95\% fewer \gls{gflops}, while achieving similar performance.

An ablation study was conducted to quantify the contribution of each architectural component. The results indicate that the MergeBlocks, dual-path design, and multi-scale connections are among the main components contributing to the model's segmentation performance and efficiency. The source code is publicly available at:
\url{https://github.com/klingerkrieg/UMergeNet}.

COMMITTEE MEMBERS:
Presidente - 2177445 - BRUNO MOTTA DE CARVALHO
Interna - 1350250 - ANNE MAGALY DE PAULA CANUTO
Externo à Instituição - FÁBIO MANOEL FRANÇA LOBATO - USP