Automated Lesion Segmentation in Medical Imaging via Integration of nnU-Net Optimization and SAM Approach

Deep learning has transformed medical imaging by enabling earlier and more accurate disease diagnosis. Lesion and tumor segmentation, essential for analyzing and tracking morphological changes, is commonly done with U-Net variants, though these often lack cross-domain generalization and do not fully leverage foundation models like the Segment Anything Model (SAM), which still requires manual intervention to define the region of interest (ROI).

To enhance generalization and reduce manual intervention by combining the automatic optimization of nnU-Net with the precision of SAM.

Experimental evaluation of a hybrid segmentation framework for lung nodule analysis.

We propose a novel approach integrating the automatic optimization capabilities of nnU-Net for lesion detection with the high-precision segmentation of SAM, eliminating the need for manual intervention by the clinician. The method was evaluated on the LIDC-IDRI dataset, a widely recognized benchmark for lung nodule segmentation.

Our approach produces more anatomically coherent segmentations than nnU-Net alone. In many cases, the resulting boundaries more closely reflect true nodule morphology than individual expert annotations, despite high inter-expert variability.

The proposed integration of nnU-Net with SAM enables fully automated lesion segmentation without manual intervention. The method improves generalization and accuracy across medical imaging domains, achieving expert-level performance in pulmonary nodule segmentation.