Research On Small Object Detection And Segmentation Algorithms In Medical Image Based On Attention-Guided Feature Fusion

Medical image object detection and segmentation are two important steps in computer-aided diagnosis system,which can provide doctors with quantitative diagnosis basis and improve the efficiency.Recently,deep learning-based models have achieved excellent results in many medical image detection and segmentation tasks.However,due to the non-obvious features and high inter-class similarity of some small-sized targets,which can easily lead to model misclassification,the accurate detection and segmentation of small targets is still a challenging task.The multi-scale feature fusion strategy alleviates the problem of loss of small-size target features by fusing deep features with rich semantic information and shallow features with rich spatial information,and improves the detection and segmentation performance of small-size targets.However,the features of different scales are not completely aligned in the spatial position.In addition,there is also a semantic gap between semantically distinct features.The simple pre-defined feature fusion method cannot cope with the above situation well,which limits the efficiency of feature fusion and hinders further improvement of the model performance.To alleviate the above problems,this paper takes small microcalcification clusters in digital breast tomosynthesis(DBT)and tiny bacteria in sputum smear images as the entry point,and carries out research related to new feature fusion strategies,exploring the potential relationships between features by various attention mechanisms and guiding the fusion of semantically distinct features accordingly,thereby improving the efficiency of feature utilization to achieve accurate detection and segmentation of small targets in medical images.Based on the idea of attention-guided feature fusion,this paper proposes an efficient detection method for microcalcification clusters in DBT and an efficient segmentation network for bacteria in sputum smears.The main work is summarised as follows:1.To address the problem that low-contrast small microcalcification clusters has less obvious features and are easily missed by existing detection models,this paper proposes an automatic microcalcification cluster detection algorithm that combines 2D slice-by-slice detection and 3D inter-layer result fusion.In the process of 2D slice-by-slice detection,a Context Attention Pyramid Network(CAPNet)is designed to mine and focus on the regions with insufficient features in the shallow feature map by calculating the uncertainty,which effectively improves the efficiency of deep features to complement shallow features and increases the detection rate of microcalcification clusters.In addition,a microcalcification response branch is designed,and a Difference of Gaussian(DOG)filter is used to provide more reliable positive and negative samples for this branch during the training phase,improving the model’s ability to distinguish microcalcifications from surrounding breast tissue and reducing false positives.The performance of the proposed method is validated with a real clinical dataset,and the experimental results show that CAPNet have a detection sensitivity of 91.56%at 1 false positive per DBT volume(FP/volume)and 93.51%at 2 FPs/volume,outperforming other representative detection models.2.To address the problems of small bacterial size,low contrast and similar morphology of different types of bacteria,which easily lead to mis-segmentation of existing models,a dual-branch deformable cross-attention fusion network(DB-DCAFN)is proposed,which uses convolution and self-attention in parallel to simultaneously extract multi-level local features and global features to enhance the model’s ability to determine the type of bacteria and locate the boundary of bacteria.To bridge the semantic gap between local and global features,a deformable cross-attention(DCA)module is designed in DB-DCAFN.The module models the potential relationships between local and global features,and performs deformable feature sampling through the learned relationships to achieve sparse and efficient cross-attention,thus guiding the efficient fusion of the two features.In addition,a feature assignment fusion(FAF)module is designed to enhance the useful features during skip-connection by an adaptive weighting strategy,further facilitating accurate segmentation of bacteria.The effectiveness of the proposed network is validated with a real clinical sputum smear dataset comprising three bacterial types:Acinetobacter baumannii(Aba),Klebsiella pneumoniae(Kpn)and Pseudomonas aeruginosa(Pae).The experimental results show that DB-DCAFN outperforms other mainstream segmentation networks and can effectively segment bacteria from sputum smears.