Detection of pulmonary nodule lesions based on improved Mask R-CNN algorithm

Lung cancer primarily develops from malignant nodules in the lungs. Early screening of lung nodules is crucial for detecting early-stage lung cancer. However, early lung nodules often lack distinct features, and the presence of non-pulmonary parenchymal areas and lung soft tissues significantly interferes with their differentiation and detection. To address the characteristics of pulmonary CT images, an improved Mask R-CNN lung nodule detection model is proposed, incorporating deep learning techniques. A hybrid attention mechanism is introduced into the feature backbone network to establish spatial and channel relationships within the feature maps, thereby enhancing the model's ability to extract features from the target and improving the correlation of shallow feature information. Additionally, the anchor box sizes are modified to five new dimensions to better capture the features of lung nodules of varying sizes during the detection process. Experimental validation of the improved algorithm demonstrates a detection sensitivity of 89.7%, an average segmentation similarity coefficient of 96%, and an average precision of 88.1%.