3D-PulCNN: Pulmonary cancer classification from hyperspectral images using convolution combination unit based CNN.

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Pulmonary cancer classification is crucial for diagnosis, but current methods using color images face accuracy challenges.
• Hyperspectral imaging offers rich spatial and spectral information, potentially improving classification accuracy.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for pulmonary cancer subtype classification using hyperspectral images.
• To leverage both spatial and spectral information for enhanced diagnostic accuracy.

Main Methods:

• A convolution combination unit (CCU)-based 3D convolutional neural network (3D-PulCNN) was proposed, fusing features from multiple convolution scales.
• The model utilizes reduced network parameters and complexity compared to VGGNet.
• 3D-UNet was employed for cancer cell segmentation to analyze morphological characteristics.

Main Results:

• The 3D-PulCNN model achieved an overall average accuracy (OA) of 0.962.
• Precision, Recall, and Kappa values exceeded 0.920, demonstrating superior performance over 2D-VGGNet.
• Quantitative data from segmented cancer cell morphology were obtained for classification explanation and prognosis.

Conclusions:

• The proposed 3D-PulCNN effectively classifies pulmonary cancer subtypes from hyperspectral images by integrating spatial and spectral data.
• This approach offers a significant advancement over traditional methods, paving the way for more accurate lung cancer diagnosis and prognosis.
• Integration with 3D-UNet provides quantitative morphological data, enhancing the interpretability and clinical utility of the classification results.