Study of intracranial haematoma localisation based on improved RetinaNet

Abstract

Intracranial haemorrhage is described as bleeding within the skull. It is a serious cranio-cerebral disorder recognized for its high mortality and lethality rate, which usually requires urgent follow-up diagnosis and determination of the location and subtype of intracranial hemorrhagic lesions.In this study, we experimented with multiple available deep learning architectures to localize the location of hemorrhagic lesions after traumatic brain injury (ICH). To improve the probability of successful patient resuscitation. In this paper, we propose an improved model based on RetinaNet. The accuracy problem of lesion localisation is not effeactively addressed due to the complex structure of the lesion location in intracranial haemorrhage and the large variation in the morphology of the lesion for different subtypes. To address these problems, the paper then proceeds to optimise the original RetinaNet model in terms of its feature extraction network structure, training techniques and Anchor settings. Through comparison experiments, it can be found that the improved model is better than the three target detection models, Faster R-CNN, RetinaNet and YOLOv4.