Fuente: PubMed "swarm" Sci Rep. 2025 Nov 28;15(1):42727. doi: 10.1038/s41598-025-26768-6.ABSTRACTEndodontic illnesses affect around 52% of the global population and are projected to rise by 4% by 2030. Endodontic problems necessitate precise classification for treatment planning and clinical decision-making. Radiographs and expert analysis are essential to conventional diagnostic techniques; hence, automated, data-driven alternatives are required to enhance accuracy and efficiency. A modified Swin transformer with non-overlapping local windows with shifted windows in alternating layers is employed to achieve hierarchical attention mechanisms. The fixed segments of the input image are linearly embedded and converted into token representations to execute self-attention using projected query, key, and value matrices, while shifting windows and hierarchical merging facilitate extensive contextual learning. The ultimate features are subjected to global average pooling and classification via a learnable head, constituting the essence of the proposed MSViT architecture. The selection of hyperparameters for the architecture and feature selection is executed using a hybrid approach of chaotic particle swarm optimization (CPSO) and sequential quadratic programming (SQP). The proposed model is trained and assessed using an improved version of the publicly available dataset known as the root canal dataset, which includes seven categories of endodontic disorders. The suggested model, optimized using CPSO-SQP, attains an average classification accuracy of 97.72%, a mean fitness value of 2.372 × 10-11, a mean precision of 0.9749, and a mean square error of 8.2 × 10- 04, with a computational time of 867.59 s at a learning rate of 0.0001. The suggested farmwork is evaluated against pre-trained models such as ResNet-101, VGGNet-19, InceptionV3, and EfficientNet-b0, in addition to a baseline strategy based on SA-SQP, GA-SQP, and other documented results. The proposed architecture enables to learn fine-grained anatomical and pathological features, essential for distinguishing endodontic diseases in an efficient and accurate manner.PMID:41315402 | PMC:PMC12663471 | DOI:10.1038/s41598-025-26768-6