| Intracranial hemorrhage(ICH)is the bleeding in brain parenchyma or surrounding meningeal space caused by the rupture of intracranial blood vessels,and timely diagnosis and treatment are important to improve the prognosis.Computed tomography(CT)scan is widely used for the early diagnosis of intracranial hemorrhage because of its rapid,non-invasive imaging and high sensitivity to hemorrhage.Different subtypes of intracranial hemorrhage require different interventions;therefore,the subtype of hemorrhage needs to be identified first in the diagnosis.In addition,the volume of hemorrhage is also an important factor to be considered in the prognostic evaluation.Currently,in clinical practice,the ABC/2 method is used to estimate the hematoma volume,which is prone to result in large errors,while manual delineation by physicians slice by slice is quite laborious and time-consuming.Therefore,a fast and automated method for intracranial hemorrhage volume measurement is needed.The aim of this study is to explore the application of multi-instance learning methods in intracranial hemorrhage computer-assisted diagnosis algorithms,which consists of two parts:(1)multi-instance learning based intracranial hemorrhage subtype classification;(2)multi-instance learning based weakly supervised intracranial hemorrhage segmentation.In this study,the intracranial hemorrhage dataset published by the Radiological Society of North America(RSNA)was first collated and preprocessed for the subsequent evaluation of the experimental results.There could be correlations and dependencies of hemorrhage between different slices,whereas this property has been ignored by previous multi-instance classification algorithms.Hence,in the first part,this work proposes a Transformer-based multi-instance learning network for intracranial hemorrhage subtype classification by considering the inter-layer correlation of intracranial hemorrhage.First,a Transformer-based feature fusion module is designed to extract the inter-slice correlation of CT using a self-attentive mechanism which completes the weighted fusion from instance features to global features;Second,a multi-scale integration structure is used to adapt to the variation of hemorrhage size;Finally,Focal loss is employed to effectively eliminate the impact of inter-class imbalance effect on the network.The proposed network in this study achieved an average accuracy of 0.921 for multi-label classification and an F1 score of 0.805 in 5-fold cross-validation,which has a good classification performance.In the second part,this work further generalized the multi-instance learning method to the intracranial hemorrhage segmentation problem.Existing weakly supervised segmentation algorithms have low accuracy and a high probability of false positives due to factors such as irregular shape of hemorrhage,poor edge contrast,and interference of CT artifacts.This work proposes a multi-instance learning based weakly supervised segmentation network,consisting of a multi-branch structure with a shared encoder,two independent decoders,which simultaneously accomplishes the multi-scale heatmap generation,pseudo-label extraction and optimization,and the segmentation network training process,which not only enhances the resistance ability to false positives,but also leverages the multi-branch collaboration effect to improve segmentation results.The network achieved better segmentation performance than several existing methods,with a Dice similarity coefficient of 0.822 and a volume similarity of 0.896,which is more consistent with manually outlined labels.In conclusion,this work realizes the multi-instance learning based algorithm for the subtype classification and segmentation of intracranial hemorrhage,providing important reference information for clinical evaluation such as the hemorrhage subtypes,locations and volume,which can promote the treatment efficiency and improve the medical decision-making procedures in primary hospitals and remote areas. |
PDF Full Text Request