Automated detection of hemorrhage and infarction on diffusion-weighted mri: comparative performance of yolov8, faster r-cnn, and radiologists

Anas, Tharek and Noralam, Noor Hayatul Al-akmal and Shamsuddin Perisamy, Rajeev and Ismail, Luthffi Idzhar and Hui, Soo Tze and Muda, Ahmad Sobri (2025) Automated detection of hemorrhage and infarction on diffusion-weighted mri: comparative performance of yolov8, faster r-cnn, and radiologists. Xi'an Shiyou Daxue Xuebao (Ziran Kexue Ban)/ Journal of Xi'an Shiyou University, Natural Sciences Edition, 68 (10). pp. 36-52. ISSN 1673-064X

PDF - Published Version Download (753kB)

Abstract

Background: Rapid differentiation between ischemic and hemorrhagic stroke is critical for timely treatment, yet diffusion-weighted imaging (DWI) alone poses diagnostic challenges for hemorrhage detection. Artificial intelligence (AI) offers potential to improve radiologist interpretation, but comparative evaluations of state- of-the-art object detection models on stroke MRI remain limited. Objective: To evaluate and compare the performance of YOLOv8 and Faster R-CNN for automated detection of intracranial hemorrhage and acute infarction on DWI, benchmarked against expert neuroradiologists. Methods: In this retrospective single- center study, 1,000 adult DWI cases were analyzed, comprising 334 hemorrhage, 333 infarct, and 333 normal studies. Images were annotated by neuroradiologists, and models were trained with and without augmentation. Performance was assessed at lesion and image levels using precision, recall, mean average precision (map), confusion matrices, and inference time. Binary hemorrhage detection was compared with radiologists using McNemar’s test. Results: YOLOv8 achieved higher recall and map than Faster R-CNN, particularly for small infarcts and subtle hemorrhages. With augmentation, recall improved to 0.886 and mAP@0.5 reached 0.903. Binary hemorrhage detection yielded sensitivity 0.91, specificity 0.88, and accuracy 0.90. Radiologists achieved near-perfect accuracy of 0.99, while Faster R-CNN lagged with sensitivity 0.82. YOLOv8 processed each image in <15 MS, compared to >40 MS for Faster R-CNN. Conclusion: YOLOv8 demonstrated superior accuracy and efficiency compared with Faster R-CNN, approaching radiologist-level sensitivity. These findings support the potential of one-stage detectors to augment radiologists in real-time stroke workflows, warranting further multicenter and multi-sequence validation.

Item Type:	Article (Journal)
Subjects:	R Medicine > R Medicine (General) R Medicine > RC Internal medicine
Kulliyyahs/Centres/Divisions/Institutes (Can select more than one option. Press CONTROL button):	UNSPECIFIED
Depositing User:	Dr Rajeev Shamsuddin Perisamy
Date Deposited:	26 Nov 2025 16:54
Last Modified:	26 Nov 2025 16:54
Queue Number:	2025-11-Q146
URI:	http://irep.iium.edu.my/id/eprint/124440

Actions (login required)

View Item