Related Experiment Video
Updated: May 13, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
Automated Lesion Segmentation in Medical Imaging via Integration of nnU-Net Optimization and SAM Approach
Alejandro Jerónimo1, Ignacio Rojas1, Olga Valenzuela2
1Computer Engineering, Automatics and Robotics Department, University of Granada, Spain.
Biomedical Engineering and Computational Biology
|May 12, 2026
Summary
This study introduces a novel deep learning method combining nnU-Net and Segment Anything Model (SAM) for automated lung nodule segmentation. The hybrid approach enhances generalization and accuracy, achieving expert-level performance without manual intervention.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Disease Diagnosis
- Computational Pathology
Background:
- Deep learning, particularly U-Net variants, is crucial for medical image analysis, yet often struggles with cross-domain generalization and requires manual input.
- Foundation models like the Segment Anything Model (SAM) offer advanced segmentation but necessitate manual definition of regions of interest (ROIs).
- Current methods for lesion and tumor segmentation lack sufficient generalization and automation, hindering clinical application.
Purpose of the Study:
- To develop a hybrid segmentation framework integrating nnU-Net and SAM for enhanced generalization and reduced manual intervention in medical imaging.
- To achieve fully automated lesion segmentation, eliminating the need for clinician input.
- To improve the accuracy and anatomical coherence of lung nodule segmentation.
Main Methods:
- A novel approach combining the automatic optimization of nnU-Net for lesion detection with the high-precision segmentation of SAM.
- Integration of nnU-Net's capabilities with SAM to eliminate manual intervention for region of interest definition.
- Evaluation of the hybrid framework on the LIDC-IDRI dataset, a benchmark for lung nodule segmentation.
Main Results:
- The hybrid approach yields more anatomically coherent segmentations compared to nnU-Net alone.
- Segmentation boundaries generated by the proposed method more accurately reflect true nodule morphology.
- The results demonstrate improved performance despite high inter-expert variability in annotations.
Conclusions:
- The integration of nnU-Net with SAM enables fully automated lesion segmentation, removing the need for manual input.
- The method demonstrates improved generalization and accuracy across medical imaging domains.
- The proposed framework achieves expert-level performance in pulmonary nodule segmentation.