Key Technical Research Of Convolutional Neural Networks In The Identification Of Typical Oral Diseases

In November 2022,the World Health Organization released its latest “Global Oral Health Report” revealing that nearly half of the global population—approximately 3.5 billion individuals—are affected by oral diseases,with the incidence rate increasing faster than population growth.Diseases with high prevalence such as dental caries,periapical periodontitis,and malocclusion,along with life-threatening conditions like oral cancer,have received significant attention.Timely and accurate screening of these diseases is critical for safeguarding public health.Traditional oral disease diagnosis relies largely on subjective judgment and clinical experience of healthcare providers,which often leads to misdiagnosis and missed diagnosis due to the concentration and scarcity of high-quality medical resources.To advance the development of intelligent oral healthcare and improve the efficiency and accuracy of disease diagnosis by healthcare professionals,this study delves into the application research of Convolutional Neural Network(CNN)in the identification of typical oral diseases.In the context of clinical oral diagnosis,the four most auxiliary imaging modalities are commonly employed: panoramic radiographs,cephalometric radiographs,periapical radiographs,and histopathological images.Within each modality,this study focuses on identifying and analyzing representative types of diseases,including maxillary first molar ectopic eruption related to malocclusion,adenoid hypertrophy,dental caries and periapical periodontitis that can degrade quality of life by causing tooth pain,as well as life-threatening oral squamous cell carcinoma.Several technical challenges arise when applying CNNs to the field of oral medicine.These include efficient data collection and annotation,label noise filtering,precise identification of disease states,and model optimization to adapt to resource-constrained environments.The innovative contributions and primary focus of this study encompass the following four aspects:1.Efficient Data Annotation and Label Noise FilteringThis thesis addresses three core challenges associated with constructing a dental image dataset-sample collection and annotation,high-value sample selection,and label noise filtering.Firstly,a rigorous and scientific sample collection and annotation method was established through the integrated application of statistical process control,collective intelligence,and expert consensus,which not only enhanced the representativeness and diversity of the dataset but also minimized subjective biases in the annotation process,thereby improving the quality and consistency of label data.Secondly,an active learning-based strategy was proposed for effective selection of high-value samples,thus improving annotation resource utilization.Lastly,to optimize the performance of the CNN model,an improved representation learning-based label noise filtering method was introduced to reduce label noise in the dataset and enhance model accuracy and robustness.2.Single-Label Oral Disease RecognitionThis thesis mainly investigates the single-label disease recognition problem in dental images,treating it as a small-sample supervised learning task due to limited annotated samples.Initially,an enhanced pixel-space feature augmentation method based on the Mixup algorithm was proposed,specifically optimized for dental images,and incorporated a Beta distribution mixup to generate more reasonable augmented samples,thereby reducing model prediction uncertainty in small or sparse sample scenarios.Subsequently,a Mixup-based method was applied in the feature space by mixing coarse-segmented images of lesion areas to increase feature diversity during training.Lastly,a series of human-machine comparative experiments were designed to evaluate the effectiveness and applicability of the proposed methods in real clinical settings.The results demonstrated that the methods could improve the work efficiency,diagnostic consistency,and accuracy of dental practitioners.3.Multi-Label Oral Disease RecognitionThis thesis delves into the single-image multi-label disease recognition problem based on CNN models,with caries and periapical periodontitis recognition tasks in dental images as specific application scenarios.A model structure for multi-label disease recognition based on adaptive subspace was proposed.This structure builds two independent mapping subspaces in the feature space,which are then precisely predicted by a shared classifier.An adaptive subspace attention model structure was designed,which dynamically constrains feature expressions within the subspaces,further enhancing the discriminative power and accuracy of the model.Finally,through a series of clinical comparative experiments,the practicality and effectiveness of the proposed methods in real medical scenarios were validated.4.CNN Model Compression and Optimization for Resource-Constrained Computing DevicesThis thesis investigates the adaptation of CNN in resource-constrained environments.It introduces a dependency-based parameter organization technique and a joint importance evaluation method for parameters,achieving more refined pruning and higher compression ratios.Additionally,by employing reparameterization techniques,the study enhances the model’s runtime efficiency during the inference phase,making it more suitable for resourcelimited computing devices.The improvements in inference efficiency are empirically validated through tests on both CPU and GPU platforms.In conclusion,the model proposed in this thesis has been successfully integrated into an intelligent oral medicine auxiliary diagnostic system.This system has been deployed in realworld settings,including West China Hospital of Stomatology,Sichuan University,Department of Stomatology,the First People’s Hospital of Ziyang and several primary-level oral clinics.Preliminary clinical evaluations indicate that the system has had a positive impact on reducing clinicians’ workload and improving diagnostic efficiency.Notably,healthcare facilities employing this system have reported a significant increase in work efficiency.Future work will focus on elucidating the interpretability of deep learning models,encompassing both mathematical and biological aspects of interpretability.