<![CDATA[Acute ischemic stroke must be treated within a few hours, and the first scan is usually a non-contrast CT (NCCT), on which an early infarction is often hard to see. Automatic segmentation assist image interpretation, and multi-scale convolution is a commonly proposed approach. This study measured how much such a block improves U-Net performance for this task, and what it costs. Every convolutional block of U-Net, in the encoder and the decoder, was replaced by a block with three parallel branches using 1×1, 3×3 and 5×5 kernels. Both networks were trained on the public Acute Ischemic Stroke Dataset, on 375 studies with supratentorial infarcts. The data were divided by study; the scans of the dataset’s own test split were kept for testing, and augmentation was applied to the training subset alone. Four configurations were compared: each architecture with and without augmentation. On the 51 test patients, Multi-scale U-Net did not improve overlap. Its Dice score was 0.6372 against 0.6632 for U-Net with augmentation, and the paired difference per patient was 0.0233 in favour of U-Net (95 % CI −0.060 to 0.013, p = 0.19). The multi-scale block instead altered the balance between the two types error: sensitivity increased by 0.060 (p = 0.002), and precision decreased by 0.100 (p < 0.001). Augmentation improved both networks (+0.106 and +0.048 Dice per patient, p < 0.001), mainly by cutting false-positive lesions from 15.5 to 5.6 per patient. The multi-scale block costs 2.31 times as many parameters and 3.37 times as much inference time. An ablation over the target definition favoured the dataset authors' labels: adding the label for infarct invisible on NCCT cost 0.040 Dice and tripled false-positive lesions. Both networks segment small infarcts poorly, with Dice below 0.16 under 100 pixels, which remains the main obstacle.]]>
