In view of the problems of vascular pleomorphism on transverse sections and sampling imbalance in the process of detection， an improved Libra Region-Convolutional Neural Network （R-CNN） cerebral arterial stenosis detection algorithm was proposed to detect internal carotid artery and vertebral artery stenosis in Computed Tomography Angiography （CTA） images. Firstly， ResNet50 was used as the backbone network in Libra R-CNN， Deformable Convolutional Network （DCN） was introduced into the 3， 4， 5 stages of backbone network， and the offsets were learnt to extract the morphological features of blood vessels on different transverse sections. Secondly， the feature maps extracted from the backbone network were input into Balanced Feature Pyramid （BFP） with the Non-local Neural Network （Non-local NN） introduced for deeper feature fusion. Finally， the fused feature maps were input to the cascade detector， and the final detection result was optimized by increasing the Intersection-over-Union （IoU） threshold. Experimental results show that compared with Libra R-CNN algorithm， the improved Libra R-CNN detection algorithm increases 4.3， 1.3， 6.9 and 4.0 percentage points respectively in AP， AP₅₀， AP₇₅ and AP_S， respectively_{on the cerebral artery CTA dataset； on the public CT dataset of colon polyps， the improved Libra R-CNN detection algorithm has the AP， AP50， AP75 and APS increased by 6.6， 3.6， 13.0 and 6.4 percentage points， respectively. By adding DCN， Non-local NN and cascade detector to the backbone network of Libra R-CNN algorithm， the features are further fused to learn the semantic information of cerebral artery structure and make the results of narrow area detection more accurate， and the improved algorithm has the ability of generalization in different detection tasks.}

The detection precision is low when the diatom training sample size is small， so a Multi-scale Multi-head Self-attention （MMS） and Online Hard Example Mining （OHEM） based few-shot diatom detection model， namely MMSOFDD was proposed based on the few-shot object detection model Two-stage Fine-tuning Approach （TFA）. Firstly， a Transformer-based feature extraction network Bottleneck Transformer Network-101 （BoTNet-101） was constructed by combining ResNet-101 with a multi-head self-attention mechanism to make full use of the local and global information of diatom images. Then， multi-head self-attention was improved to MMS， which eliminated the limitation of processing single object scale of the original multi-head self-attention. Finally， OHEM was introduced to the model predictor， and the diatoms were identified and localized. Ablation and comparison experiments between the proposed model and other few-shot object detection models were conducted on a self-constructed diatom dataset. Experiment results show that the mean Average Precision （mAP） of MMSOFDD is 69.60%， which is improved by 5.89 percentage points compared with 63.71% of TFA； and compared with 61.60% and 60.90% the few-shot object detection models Meta R-CNN and Few-Shot In Wild （FSIW）， the proposed model has the mAP improved by 8.00 percentage points and 8.70 percentage points respectively. Moreover， MMSOFDD can effectively improve the detection precision of the detection model for diatoms with small size of diatom training samples.