DeepSEED: 3D Squeeze-and-Excitation Encoder-Decoder Convolutional Neural Networks for Pulmonary Nodule Detection

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology and Oncology

Background:

• Pulmonary nodule detection is crucial for lung cancer screening using low-dose computed tomography (CT).
• Deep learning models face challenges in generalization due to imbalanced positive and negative samples in nodule detection.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced pulmonary nodule detection.
• To overcome sample imbalance and reduce false positives in lung nodule identification.

Main Methods:

• A novel deep 3D convolutional neural network with an Encoder-Decoder structure and a region proposal network was developed.
• A dynamically scaled cross entropy loss function was utilized to address sample imbalance and reduce false positive rates.
• Squeeze-and-excitation structures were incorporated for effective feature learning and inter-dependency analysis of feature maps.

Main Results:

• The proposed method demonstrated superior performance compared to existing state-of-the-art nodule detection techniques.
• Validation on LIDC/IDRI and LUNA16 datasets confirmed the method's effectiveness.
• Ablation studies validated the contribution of the proposed components.

Conclusions:

• The novel deep learning approach effectively improves pulmonary nodule detection accuracy and generalization.
• The method offers a significant advancement for lung cancer screening through more reliable nodule identification.