CoLe-CNN: Context-learning convolutional neural network with adaptive loss function for lung nodule segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate lung nodule segmentation in computed tomography (CT) is vital for tumor characterization.
• Manual segmentation is time-consuming and hinders clinical practice.
• A novel Convolutional Neural Network (CNN) is proposed to address these challenges.

Purpose of the Study:

• To develop an efficient CNN for accurate lung nodule segmentation.
• To introduce an innovative loss function and segmentation strategy.
• To compare the network's performance against the state-of-the-art and human radiologists.

Main Methods:

• A novel CNN architecture learns nodule context by generating background and secondary element masks.
• Nodule detection is achieved by subtracting the context mask from the original CT scan.
• An asymmetric loss function compensates for annotation errors; trained and tested on the LIDC-IDRI database.

Main Results:

• The proposed method demonstrates performance comparable to human radiologists.
• Segmentation masks are nearly indistinguishable from those created by expert radiologists.
• Outperforms state-of-the-art methods with improved F1 score (3.3%) and IoU (4.7%) in CT nodule segmentation.

Conclusions:

• The CNN combines UNet properties with Multi Convolutional Layers for enhanced pattern recognition.
• The method improves nodule border detail, even in noisy conditions.
• Applicable for single CT slice segmentation, serving as a foundation for future 3D segmentation software.