SwiftMedSAM: An Ultra-Lightweight Prompt-based Universal Medical Image Segmentation Model for Highly Constrained Environments

Youngbin Kong, Kwangtai Kim, Seoi Jeong, Kyu Eun Lee, HyounJoong Kong

Keywords: Medical Imaging Segmentation, Segment Anything, Deep Learning

Abstract:

Medical image segmentation is a crucial step for accurate
diagnosis and treatment planning, as it provides quantitative informa-
tion about anatomical structures and pathological lesions in various
clinical scenarios. However, the existing methodologies have limitations
in terms of their generalizability and computational efficiency. In this
study, we propose SwiftMedSAM, an ultra-lightweight prompt-based
general model, to enable efficient medical image segmentation even in
resource-constrained environments. Based on LiteMedSAM, we signifi-
cantly reduced the model size and computational complexity through the
hyperparameter optimization of the image encoder and mask decoder
components. The developed model shows remarkable segmentation per-
formance across various imaging modalities and anatomical structures
while enabling real-time inference in resource-limited computing envi-
ronments. The experimental results demonstrate that SwiftMedSAM
outperforms the existing methodologies in terms of the trade-off between
accuracy and efficiency. SwiftMedSAM achieved a validation score of 0.75
on the validation dataset. Owing to its unprecedented generalizability
and low computational cost, SwiftMedSAM is expected to enable high-
quality medical image analysis in resource-constrained settings, thereby
contributing to advancements in precision medicine and telemedicine.

Submission Number: 4