Speaker
Description
Ischemic stroke is a leading cause of neurological disability, and early detection of its lesions on non-contrast CT (NCCT) images remains challenging due to their subtle appearance. This study aims to analyze the localization of ischemic stroke lesions on NCCT images and evaluate its performance using the integration of a Residual Network and a Convolutional Block Attention Module (ResNet-CBAM). The AISD dataset comprising 2,380 slices from 388 patients was used with a patient-level split. Preprocessing included HU windowing (level 35, width 35), resizing to 224×224, ImageNet normalization, and ±5% translation augmentation. The ResNet50 architecture was modified with AdaptiveAvgPool2d(2) and a regression head for bounding box regression, with three variants: Baseline, CBAM Layer4, and CBAM All Block. Models were trained using Smooth L1 Loss, the Adam optimizer, and early stopping. Evaluation used Mean IoU, Center Localization Error (CLE), Generalized IoU (GIoU), and a paired t-test. Results on 336 test slices showed that CBAM All Block achieved the highest Mean IoU (0.1426) with 14.58% of predictions exceeding IoU > 0.5, followed by CBAM Layer4 (0.1339) and Baseline (0.1265). Under a single paired test, the difference between CBAM All Block and Baseline showed p = 0.0211 (p < 0.05), but after Bonferroni correction, this difference was no longer significant. The lowest CLE was attained by CBAM Layer4 (29.50 pixels), while GIoU values across the three models were practically equivalent (−0.2625 to −0.2727). Thus, no single model consistently outperformed across all evaluation metrics. The low absolute performance reflects the difficulty of localizing subtle early ischemic lesions on NCCT, particularly for small lesions.
Keywords: ischemic stroke; NCCT; localization; ResNet50; CBAM; bounding box; deep learning.
