Grid Inference for Interactive Segmentation of Lung Tumors in CT with Deep Learning

Accurate delineation of lung lesions in computed tomography (CT) is crucial for tumor assessment, treatment planning, and quantitative analysis. However, manual segmentation is time-consuming, particularly in patients with multiple lesions, whereas automated methods may require substantial manual correction. Among interactive segmentation approaches, click-based methods provide a practical way for clinicians or trained users to guide lesión delineation through a limited number of interactions. Minimizing user interactions while preserving accurate lesion boundaries is therefore essential for efficient clinical workflows. We propose an interactive segmentation framework that combines positive clicks within the lesion, negative clicks on the background, and block-wise inference. We adapt “Efficient Mask Correction for Click-Based Interactive Image Segmentation”, originally developed for natural RGB images, to axial CT slices. We add a convolutional neural network after the final layer of the original model, with only the newly introduced parameters trained while the original model remained frozen. Each axial slice was divided into square blocks, processed independently, and reassembled into the complete segmentation. The model achieved a Dice similarity coefficient (DSC) of 0.96 on Medical Segmentation Decathlon (MSD) and 0.95 on Hospital Clínico Universidad de Chile (HCUCH) datasets, outperforming interactive models such as Mem3D (0.82) and RiRL (0.89) on MSD, and the automatic 3D nnU-Net (0.85) on HCUCH, for lung tumor segmentation. The average number of clicks required to reach 80%, 85%, and 90% DSC in one block was 2.58, 3.14, and 4.55 on MSD, and 2.18, 2.77, and 4.08 on HCUCH, respectively. The results demostrate that the proposed strategy achieves high segmentation accuracy with a small number of clicks. By allowing users to interact with the complete CT image while block-wise inference operates in the background, the framework has the potential to support efficient and reproducible tumor delineation, reducing annotation effort and facilitating quantitative assessment and treatment planning.