Objective To construct diagnostic models based on YOLOv8n for qualitative diagnosis of common intracranial diseases and localization diagnosis of cerebral hemorrhage on cranial CT, and to evaluate its diagnostic performance and clinical application value, so as to provide a lightweight auxiliary tool for rapid and accurate diagnosis of intracranial diseases. Methods Plain cranial CT images from the First Affiliated Hospital of Guilin Medical University were retrospectively collected to construct 2 independent datasets: (1) qualitative diagnosis dataset: 1 000 images each of cerebral infarction, cerebral hemorrhage, pituitary adenoma, and meningioma were enrolled, totaling 4 000 images; (2) localization diagnosis dataset: 500 images of cerebral hemorrhage at each of 9 anatomical sites were enrolled, totaling 4 500 images. Lesion boundaries were annotated via Labelimg software, and after verification by radiologists, each independent dataset was randomly divided into training, validation, and test sets at a ratio of 8∶1∶1. Based on the lightweight YOLOv8n architecture, diagnostic models were constructed, with YOLOv5n as the control model; model performance was evaluated using precision, recall, mean area under the precision-recall curves of all categories when the intersection over union threshold was set to 0.5 (mAP@0.5), and F1 score. An independent clinical validation set was used for blinded controlled validation against 3 radiologists with different levels of experience, in order to assess diagnostic accuracy, Kappa agreement, and diagnostic time. Results (1) Qualitative diagnosis model. ① Performance evaluation: precision=0.9230, recall=0.9600, mAP@0.5=0.9640, F1 score=0.9400, and overall diagnostic accuracy=0.9350. ② Model comparison: The YOLOv8n model showed overall performance similar to YOLOv5n, but with better specificity. ③ Clinical validation: The YOLOv8n model demonstrated excellent diagnostic consistency with the 3 radiologists with different levels of experience (all Kappa coefficients>0.9, all P<0.05); its accuracy of 0.9350 was higher than those of the junior radiologist's (0.8850) and mid‑level radiologist's (0.9200), but lower than that of senior radiologist's (0.9900); its total diagnostic time for 200 images was only 0.18 min, far faster than manual diagnosis (6.60-10.00 min). (2) Localization diagnosis model: ① Performance evaluation: precision=0.9580, recall=0.9800, mAP@0.5=0.9730, F1 score=0.9600. ② Clinical validation: The YOLOv8n model also showed excellent diagnostic consistency with the 3 radiologists with different levels of experience (all Kappa coefficients>0.9, all P<0.05); its accuracy of 0.9533 was higher than that of the junior radiologist's (0.9178), but lower than those of the mid‑level radiologist's (0.9644) and the senior radiologist's (0.9889); its total diagnostic time for 450 images was only 0.38 min, far faster than manual diagnosis (17.90-24.00 min). Conclusion The YOLOv8n‑based models for qualitative diagnosis of common intracranial diseases and localization diagnosis of cerebral hemorrhage on cranial CT achieve high diagnostic accuracy and inference efficiency, with good consistency with radiologists' interpretations. It can serve as an auxiliary diagnostic tool to improve diagnostic efficiency and result consistency in primary healthcare institutions and emergency settings.

无