Pulmonary Nodule Classification Using a Multiview Residual Selective Kernel Network.

Area of Science:

• Medical imaging and artificial intelligence
• Computer-aided diagnosis (CAD) for oncology

Background:

• Lung cancer is a leading cause of global mortality, necessitating early detection.
• Existing computer-aided diagnosis systems for lung nodule classification show room for improvement in accuracy.
• Deep learning strategies offer potential for enhancing diagnostic capabilities in medical imaging.

Purpose of the Study:

• To develop and investigate a novel computer-aided diagnosis scheme for predicting the malignant likelihood of lung nodules using deep learning.
• To improve the classification accuracy of malignant pulmonary nodules in computed tomography (CT) images.

Main Methods:

• An efficient residual selective kernel (RSK) block was designed to address nodule diversity.
• A multiview RSK network (MRSKNet) was established, integrating axial, coronal, and sagittal CT image planes.
• Handcrafted texture features, specifically homogeneity (HOM), were concatenated with CT intensity images to enhance the network input.

Main Results:

• The proposed MRSKNet achieved a high area under the receiver operating characteristic (AUC) of 0.9711 on the LIDC-IDRI dataset.
• The network demonstrated superior classification accuracy and a better balance between recall and specificity compared to state-of-the-art methods.
• The integration of handcrafted texture features with deep learning significantly advanced classification performance.

Conclusions:

• The developed pulmonary nodule classification framework shows great efficacy in facilitating lung cancer diagnosis.
• Combining handcrafted texture features with deep learning models is a promising approach for improving CAD systems.
• The MRSKNet architecture holds potential for advancing early lung cancer detection in clinical practice and further image processing applications.