This paper presents a deep-learning framework for the automated segmentation of acute ischemic stroke lesions on noncontrast computed tomography (NCCT). To improve the extraction of subtle low-contrast infarct features, a window-setting optimization (WSO) module was integrated into a 2D U-Net architecture. The framework was developed and evaluated using 386 cases selected from a public acute ischemic stroke dataset, with paired diffusion-weighted magnetic resonance imaging used as the reference for lesion annotation. The WSO module learned adaptive combinations of three clinically relevant CT window settings using 1 × 1 convolutions, adding only six trainable parameters. In addition, pixel-, lesion-, slice-, and patient-level evaluation schemes were adopted to provide a practical assessment of segmentation performance. Experimental results on the independent test set showed that the proposed model achieved a Dice score of 0.661, a balanced accuracy of 0.784, a sensitivity of 0.570, and a precision of 0.788. The WSO module also improved performance across multiple encoder-decoder backbones. At the lesion level, the model achieved an F1-score of 0.758 and a precision of 0.90 for moderate-to-large infarcts. These results demonstrate that the proposed framework can effectively enhance infarct segmentation on standard NCCT images without hardware modification.