MS-YOLO11: a multidimensional synergistic you only look once 11 algorithm for pulmonary nodule detection

Objective To address the misdiagnosis problem of pulmonary nodules caused by complex background interference, difficulty in detecting small targets, and morphological diversity in computed tomography (CT) imaging diagnosis, this study proposes an improved deep learning algorithm model to improve the detection accuracy and robustness of pulmonary nodules. Method Based on the You Only Look Once (YOLO)11 framework, the Multidimensional Collaborative Attention (MCA) method and the Synergistic Multi Attention Transformer (SMAT) method were combined to construct a novel model, Multidimensional Synergistic YOLO11 (MS-YOLO11). MCA dynamically enhances the representation of key features by applying attention mechanisms along the channel, height, and width dimensions. At the same time, SMAT combines pixel, channel, and spatial multi-attention mechanisms to optimize global and local feature interaction. This experiment is based on the 2016 Lung Nodal Analysis dataset (LUNA16) and evaluates the performance of the new model on segmentation and non-segmentation lung parenchyma data through ablation experiments and comparative experiments. This study mainly compared three indicators: average precision of small target ( AP small ), average recall rate of small target ( AR small ), and mean average precision ( mAP 50 ) to further verify the performance of the new model. Result The mAP 50 of MS-YOLO11 on the dataset without segmented lung parenchyma was 83.66%, which was 4.14 percentage points higher than the baseline YOLO11 (79.52%). In small object detection, AP small increased from 34.89% to 37.88% (+2.99%), and AR small increased from 46.25% to 50.74% (+4.49%). Compared to traditional models such as Single Shot MultiBox Detector (SSD) (49.07%) and Faster Region-based Convolutional Neural Network (Faster-R-CNN) (61.76%), MS-YOLO11 achieved significantly better mAP 50 while maintaining a lightweight structure of only 9.35 MB. Moreover, even when compared to recent advanced models specifically designed for small object detection, MS-YOLO11 achieved superior results across all key metrics with a smaller size, highlighting its robustness. Experiments further showed that using the entire unsegmented lung parenchyma image data preserves complete nodule features and leads to optimal detection performance. The synergistic effect of the MCA and SMAT methods effectively mitigates background interference, improves the accuracy of small object detection, and enhances overall model robustness.