Research On Key Technologies Of Medical Image Aided Diagnosis Based On Artificial Intelligence

The rapid development of artificial intelligence technology has facilitated significant breakthroughs in medical image aided diagnosis technology research,and it has quickly transitioned from the experimental stage to the clinical trial stage,alleviating the long-standing problem of scarce medical resources in our country.However,despite this,the successful deployment of medical image aided diagnosis systems faces numerous challenges due to the lack of high-quality medical datasets.These challenges include:(1)the problem of low recognition accuracy for tail class diseases due to the long-tailed distribution of medical data.The prevalence of diseases varies in clinical scenarios,resulting in highly imbalanced long-tailed distributions of medical datasets.This dominance of head diseases during the training process weakens the learning of rare and uncommon diseases.(2)The issue of the model’s inability to detect unknown diseases due to closed-set identification settings.Intelligent aided diagnosis models are mostly developed under closed-set identification settings,which fail to detect unknown diseases that may appear at any time in clinical scenarios,leading to misdiagnosis and missed detections.(3)The problem of incremental aided diagnosis for few-shot novel diseases.Unknown novel diseases constantly emerge in dynamic healthcare environments,requiring incremental aided diagnosis for the system to operate reliably.However,the data volume for new diseases is often limited,making it challenging to ensure correct decisions for these novel diseases even with retraining of the model.In response to the aforementioned challenges,this article investigates key technologies of medical image aided diagnosis based on AI.The research focuses on three main areas:long-tail distribution medical image recognition,medical image open-set recognition,and incremental auxiliary diagnosis for few-shot novel diseases.The specific research content is as follows:(1)To address the issue of low accuracy in identifying tail diseases in medical image recognition caused by long-tail distributions,a MultiBranch Network for long-tail medical image recognition technology(MBNM)is proposed.Firstly,a multi-branch network architecture is introduced,which includes a regular learning branch,a tail learning branch,and a fusion balance branch.These three branches focus on recognizing common diseases,and rare diseases,and improving the overall performance of long-tail data recognition,respectively.This approach alleviates the problem of bias in the recognition of a unified classifier.Secondly,a tail learning branch based on feature storage is proposed,which stores historical features using a reverse sampling strategy to train independent classifiers.This enhances the intra-class diversity of tail classes and improves decision-making ability for rare and uncommon diseases.Finally,a fusion balance branch based on adaptive Dice loss is proposed,which better balances the decision advantages of the regular learning branch and tail learning branch.Additionally,the adaptive Dice loss helps alleviate model bias caused by the different difficulty levels of disease diagnosis.MBNM achieved 2.78%and 3.21%improvement in balanced accuracy(BACC)on the Skin-7 dataset in skin diseases and the F-OCT dataset in ophthalmology,respectively,demonstrating the effectiveness of the MBNM model.Furthermore,the generalization and advancement of MBNM were validated on two publicly available natural image datasets,particularly on the long-tail CIFAR-100 dataset with an imbalance factor of 100,where the accuracy improved by 1.87%compared to the current state-of-the-art MBJ method.(2)In response to the problem of the model’s inability to detect unknown class diseases caused by closed-set recognition settings,a finegrained medical image open-set recognition technique based on data mixing and spatial position constraint loss(DM-SPCL)is proposed.Firstly,a simple and effective data mixing method was proposed to generate virtual unknown classes of different difficulty levels,allowing for anticipation of the fine-grained and diverse distribution of unknown classes in real clinical scenarios.Secondly,a spatial position constraint loss was introduced to control the position distribution of known classes,real unknown classes,and virtual unknown classes in the feature space.Specifically,the prototype position constraint loss forces all known classes to be distributed in the peripheral region of the feature space.At the same time,the virtual unknown classes are controlled to be distributed in the area between the known classes and the real unknown classes,thereby acting as a barrier.Furthermore,the asymmetric instance contrast loss facilitated better clustering of known class samples and pushed the virtual unknown classes away to achieve clear and concise decision boundaries between known and unknown classes.DM-SPCL achieved 4.32%and 4.56%improvement in AUROC performance on the F-OCT dataset in ophthalmology and the HyperKvasir dataset in gastrointestinal endoscopy,respectively.This demonstrates that DM-SPCL can significantly enhance the recognition performance of unknown classes while ensuring the accuracy of known class identification.Furthermore,DM-SPCL was validated for its generality and advancement on three other natural image datasets.Compared to the current state-of-the-art ARPL+CS method,DM-SPCL achieved a 1.73%improvement in AUROC performance on the CIFAR10 dataset.(3)To address the issue of incremental assistive diagnosis for FewShot new class diseases,an incremental auxiliary diagnosis technology for few-shot medical images based on Feature Augmentation and Classifier Adaptation(FSCIL-FACA)was proposed to adaptively adjust the decision boundaries between different task categories as the incremental tasks progressed.Firstly,a feature augmentation network was introduced,which improved generalization through self-supervised learning and better extraction of discriminative features unique to sparse samples through a modulation attention mechanism.Secondly,an adaptive incremental classifier was proposed to calibrate the unified classifier weights between new and old tasks using a hybrid relational projection network and adaptively adjust the feature representation to accommodate the global classification tasks.Finally,a pseudo-incremental episode training method based on meta-learning was proposed to construct pseudo-incremental tasks using pseudo-incremental episode selection methods,conduct multistage training of the hybrid relation projection network to rapidly adapt to new tasks,and extend the capability of few-shot class incremental learning.FSCIL-FACA achieved 2.79%and 1.39%improvement in average accuracy(Avg)on the HyperKvasir dataset in gastrointestinal endoscopy and the Skin-7 dataset in skin diseases,respectively,demonstrating that FSCIL-FACA significantly improves the overall performance of few-shot incremental tasks for new class diseases.The generalization of FSCILFACA was also validated on two natural image datasets where it achieved state-of-the-art results.