This paper presents a novel approach to enhance the detection and segmentation of small liver lesions in computed tomography (CT) scans using a size-focused multi-model framework. Current state-of-the-art segmentation models, primarily based on the UNet architecture, often exhibit inferior performance on small lesions due to severe class and size imbalances. We introduce a model architecture incorporating a configurable attention mechanism within the model’s skip connections and a lesion selection algorithm that compares predictions from multiple models, including a general lesion segmentation model and a small lesion-focused model, selecting the most suitable prediction. The approach was evaluated on a clinical 3-phase CT dataset and the public LiTS dataset. Results show improvements in overall lesion segmentation performance by 1.5% and 1.9% for the clinical and LiTS datasets, respectively. Additionally, the detection of small lesions improved by 4.4% and 1.8% for both datasets, respectively.