Abstract: Non-enhanced CT scanning is the first choice for the diagnosis of suspected cerebral hemorrhage in the emergency room.Medical staffs usually use CT images to manually segment the lesions of patients with suspected acute cerebral hemorrhage, and then classify them based on clinical experience.This method of manual diagnosis requires the physician's experience and is highly subjective.Moreover, the segmentation and classification tasks are performed separately, and the characteristic information associated between the two tasks cannot be fully utilized, and the time cost is high, which increases the difficulty of quickly segmenting and classifying cerebral hemorrhage lesions based on CT images.In response to the above problems, the paper proposes a model for segmentation and classification of cerebral hemorrhage images based on multi-task learning.On the one hand, the weight of the loss function is optimized according to the difficulty of learning different tasks.On the other hand, public information sharing is realized in the shallow layer of the multi-task learning network, and private information of different tasks is extracted deeply to obtain more representative features, so as to quickly and accurately segment and classify the CT images of patients with cerebral hemorrhage.The experimental results show that the segmentation annotations generated by the multi-task learning network have good visual consistency with the real annotations.Under the optimal weight, the average Dice coefficient (DSC) of all subjects is 0.828, the sensitivity is 0.842, the specificity is 0.985, and the positive predictive value (PPV) is 0.838.The accuracy, sensitivity, specificity and AUC value of multi-task learning network classification are 95.00%, 90.48%, 100.00% and 0.982, respectively.Compared with single-task deep learning, Y-Net and multi task learning assisted by classification, this method makes more effective use of relevant task information, and at the same time improves the segmentation and classification accuracy of hemorrhage lesions by adjusting the weight of the loss function.

Key words: 3DU-Net, Cerebral hematoma, Convolutional neural network, CT, Multi-task learning