Classification In Clinical Assistive Diagnosis Based On Small Sample Size Data And Feature Learning

Artificial intelligence(AI)technology has brought disruptive changes to medical models,providing a new paradigm for clinical diagnosis that relies on complex and rich medical data.Intelligent clinical auxiliary diagnosis and treatment applications have broad application scenarios.Feature learning,as its main technology implementation method,obtains more accurate and deeper knowledge from data,extracts relevant feature information to interpret data or predict specific task results,and mines the value behind data.However,medical data has low normalization,is difficult to obtain,and lacks high-quality annotated datasets,which reduces the transferability and scalability of algorithm models in new clinical tasks.Currently,small sample feature learning methods include feature optimization-based methods,data-driven enhancement-based methods,and transfer learning-based methods.However,due to their respective adaptability ranges and limitations,as well as the diversity of clinical problems and the complexity and variability of imaging and non-imaging data,there are often problems of feature redundancy,high dimensionality,model overfitting,and low efficiency.In this paper,algorithm parameter optimization and model structure improvement are carried out in three few sample feature learning directions,combined with the one-dimensional clinical analysis data and two-dimensional imaging data that clinical auxiliary diagnosis and treatment mainly rely on,and method performance testing and evaluation are realized.The specific contents are as follows:(1)For one-dimensional clinical analysis data,a novel few-sample feature learning method based on feature optimization is constructed,named SVM-MPA,which combines an advanced swarm intelligence optimization algorithm called the Marine Predatory Algorithm(MPA)and support vector machine(SVM)to achieve feature selection,SVM hyperparameter optimization,and classification at the same time.To verify the actual performance of the method,it is applied to the detection of anterior cruciate ligament injury(ACLD).Gait data of patients(35 patients and 35 normal people)are collected,and time-domain,frequency-domain,time-frequencydomain,and nonlinear features were extracted to establish a high-dimensional small-sample dataset.Then,the SVM classification error rate of 5-fold cross-validation was used to construct the fitness function of MPA,which is used to search for features and optimize two hyperparameters of SVM.By comparing with seven well-known swarm intelligence optimization algorithms,it was found that the proposed method achieved the best ACLD detection performance(sensitivity: 95.46±0.64%,specificity: 99.41±0.44%,accuracy:97.43±0.27%).In addition,post-processing by maximum voting further improved the detection performance(sensitivity: 95.50±0.89%,specificity: 99.63±0.49%,accuracy: 97.57±0.46%),which was better than the detection performance achieved by previous research.These results demonstrate the feasibility and effectiveness of the proposed method,providing a feasible solution to the one-dimensional clinical small sample data problem.(2)For two-dimensional few medical images,a data-driven feature enhancement and transfer learning approach for small sample feature learning is studied.A risk level assessment and diagnostic assistance model is established,consisting of a pre-feature segmentation network with pyramid pooling and cross-attention modules and a post-attention classification network.The pre-feature segmentation network utilized an encoding-decoding network as the backbone,which included serial double pooling,cross-attention,and pyramid modules.The serial double pooling addressed the overfitting problem caused by the complex texture of internal tissue,while the cross-attention is used for feature transfer and fusion to enhance target recognition attention.The middle layer utilized pyramid pooling modules to filter and combine compressed features at different levels,improving overall target recognition ability.The postattention classification network utilized VGG16 as the backbone and designed a new parallel double pooling attention module to enhance target recognition ability.A progressive classification method is constructed,applying feature extraction and fusion with the feature distribution map for transfer learning network fine-tuning,resulting in the optimization of classification network parameters.The method is applied to a dataset of 55 lung cancer spinal metastasis MRI images and tested using 4-fold cross-validation,achieving an average pixel accuracy of 0.815,an average target intersection over union of 0.711,and a classification accuracy of 0.95 and a classification precision of 0.89.The proposed method has excellent performance in multi-target recognition.As well as comparing the other 3 state-of-the-art methods,it has superior performance in spine risk classification.(3)A new feature enhancement small sample learning method is designed for twodimensional clinical refinement images on a data-driven basis.A novel multiple attention neural network model based on directed graph search is proposed.First,multi-attention is constructed to adaptively integrate local features and their global dependencies,simultaneously learning to focus on target structures at different scales to achieve feature region segmentation.Second,feature information and morphological parameters are introduced to improve the search algorithm of graph traversal,optimizing feature recognition and classification.The method is applied to two publicly available retinal image datasets,resulting in binary vessel segmentation.A directed graph representing the topology and spatial connections of the vessel network is constructed,utilizing local geometric information including color,diameter,and angle to decompose the complex vessel tree into multiple subtrees,ultimately classifying and labeling vessel feature points.The proposed method has been tested on the DRIVE dataset and the IOSTAR dataset containing 40 images and 30 images,respectively,with 0.863 and 0.764 F1-score of detection points and average accuracy of 0.914 and 0.854 for classification points.These results demonstrate the superiority of our proposed method outperforming state-of-theart methods in feature point detection and classification.In conclusion,we propose novel feature learning optimization algorithms applicable to different data scenarios and show excellent performance for the typical complex small sample problem in clinical aid diagnosis for classification,which has some practical application value.