Pattern classification of interstitial lung diseases from computed tomography images using a ResNet-based network...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Accurate classification of interstitial lung disease (ILD) patterns from computed tomography (CT) is crucial for patient management.
• Existing deep learning models often overlook the 3D structure of ILD patterns and utilize 2D inputs.
• High-performance ResNet-based networks like SE-ResNet and ResNeXt have not been extensively applied to ILD pattern classification.

Purpose of the Study:

• To propose and evaluate a novel deep learning model, SE-ResNeXt-SA-18, for classifying pathological patterns in ILD CT images.
• To investigate the impact of 3D input and advanced network architectures on ILD classification accuracy.
• To compare the performance of the proposed model against existing CNNs and baseline ResNet architectures.

Main Methods:

• Development of the SE-ResNeXt-SA-18 model, integrating ResNeXt's multipath design and squeeze-and-excitation network with split attention.
• Utilizing 3D convolutional neural network input for pathological pattern classification.
• Comparative analysis of SE-ResNeXt-SA-18 against ResNet-18 and SE-ResNeXt-18, evaluating the influence of input patch size.

Main Results:

• The SE-ResNeXt-SA-18 achieved superior classification performance, with average accuracy, sensitivity, and specificity of 0.991, 0.979, and 0.994, respectively, using a 32×32×16 input patch size.
• Classification accuracy increased with larger input patch sizes.
• Class activation maps indicated that high-weight regions in the SE-ResNeXt-SA-18 corresponded to specific pathological pattern features.

Conclusions:

• The SE-ResNeXt-SA-18 demonstrates superior performance in classifying ILD patterns compared to previously reported CNNs.
• The model's ability to leverage 3D information and advanced feature weighting enhances classification accuracy.
• SE-ResNeXt-SA-18 holds significant potential for accurately tracking and monitoring the progression of interstitial lung disease.