Multi-Label Softmax Networks for Pulmonary Nodule Classification Using Unbalanced and Dependent Categories.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computer-assisted diagnosis

Background:

• Current computer-assisted diagnosis systems for lung cancer struggle with multi-label classification of lung nodules.
• Existing methods using convolutional neural networks (CNNs) fail to adequately exploit statistical dependencies between labels and face challenges with imbalanced datasets.

Purpose of the Study:

• To propose MLSL-Net, a method designed to effectively perform multi-label classification of lung nodules while addressing data imbalance and label dependency.
• To optimize the performance of CNNs for lung nodule classification on imbalanced datasets.

Main Methods:

• The proposed MLSL-Net utilizes multi-label softmax loss (MLSL) to directly optimize ranking loss and AUC, reducing errors both between and within labels.
• A novel scale factor is introduced to narrow gradient discrepancies across labels and facilitate the exploitation of label dependency.

Main Results:

• MLSL-Net demonstrated effective multi-label classification performance on the LIDC-IDRI dataset and a similar dataset, even with data imbalance.
• The introduced scale factor was shown to be responsible for capturing label correlations, leading to improved prediction accuracy.

Conclusions:

• MLSL-Net offers a robust solution for multi-label lung nodule classification, successfully mitigating challenges posed by data imbalance.
• The method enhances diagnostic interpretability and accuracy in computer-assisted lung cancer detection by effectively leveraging label dependencies.