MobileLookNet: A Lightweight Convolutional Neural Network for Detection of Osseous Metastasis Using Feature Fusion and Attention Strategies

• Main Authors: Morales, Irish Danielle, Echon, Carlo Joseph, Teaño, Angelico Ruiz, Alampay, Raphael, Abu, Patricia Angela R
• Format: text
• Published: Archīum Ateneo 2024
• Subjects: bone scintigrams; convolutional neural networks; lightweight networks; metastasis; Analytical, Diagnostic and Therapeutic Techniques and Equipment; Biomedical; Computer Engineering; Electrical and Computer Engineering; Engineering; Medicine and Health Sciences
• Online Access: https://archium.ateneo.edu/intelligent-visual-env/5; https://doi.org/10.1145/3663976.3664235
• Institution: Ateneo De Manila University

This study introduces MobileLookNet, a novel lightweight architecture designed for detecting osseous metastasis in bone scintigrams on resource-constrained devices. By employing depthwise separable convolutions in parallel, utilizing inverted residuals, and integrating low-level and high-level features, MobileLookNet captures diverse levels of abstraction and extracts more individually expressive features. It outperforms traditional bone scintigraphy methods and state-of-the-art networks in metastasis detection while requiring significantly fewer floating-point operations (FLOPs) and parameters. Ablation studies reveal that feature fusion yields superior results compared to transformer-based attention strategies, highlighting the informative nature of low-level features in metastasis detection. Moreover, MobileLookNet demonstrates a trade-off between high accuracy, low FLOPs, and low parameters, where at most two can be achieved at a time. Overall, MobileLookNet shows promise in assisting nuclear medicine practitioners and enhancing metastasis detection in resource-constrained settings.