A brain-like classification method for computed tomography images based on adaptive feature matching dual-source...

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning

Background:

• Deep learning models often struggle with small datasets, leading to issues like negative transfer in transfer learning when source and target data have significant semantic differences.
• Human brains effectively filter relevant features while ignoring irrelevant ones during recognition, a mechanism not fully replicated in standard transfer learning.
• Accurate preoperative diagnosis of solitary pulmonary nodules, distinguishing between lung granuloma and lung adenocarcinoma, is crucial but challenging with limited data.

Purpose of the Study:

• To propose a novel brain-like classification method for preoperative aided diagnosis of lung granuloma and lung adenocarcinoma in patients with solitary pulmonary solid nodules using small sample sizes.
• To address the limitations of traditional transfer learning, specifically negative transfer caused by large semantic differences between source and target domains.
• To enhance the robustness and feature expression ability of deep learning models for improved diagnostic accuracy.

Main Methods:

• Developed an adaptive selected-based dual-source domain feature matching network simulating human brain feature selection to determine optimal feature matching weights between source and target networks.
• Introduced a diverse branch block in the target network to enhance feature representation by incorporating varied receptive fields and complex paths.
• Employed an ensemble classifier based on sparse Bayesian extreme learning machine for robust classification by adaptively combining base classifier outputs.

Main Results:

• The proposed method demonstrated strong performance in preoperative aided diagnosis for solitary pulmonary nodules.
• Achieved high Area Under the Curve (AUC) values of 0.9542 and 0.9356 on data from two independent centers, indicating significant diagnostic accuracy.
• The brain-inspired feature selection and ensemble classification effectively improved model robustness and diagnostic capabilities, outperforming standard approaches.

Conclusions:

• The proposed brain-like adaptive feature matching dual-source domain heterogeneous transfer learning method significantly enhances diagnostic accuracy for lung nodules in small sample scenarios.
• This approach effectively mitigates negative transfer by adaptively selecting relevant features and improving the target network's feature expression.
• The method provides a valuable and reliable diagnostic reference for clinicians in the preoperative assessment of solitary pulmonary nodules.