Comparison of Shallow and Deep Learning Methods on Classifying the Regional Pattern of Diffuse Lung Disease
Guk Bae Kim1, Kyu-Hwan Jung2, Yeha Lee2
1Biomedical Engineering Research Center, Asan Institute of Life Science, Asan Medical Center, 388-1 Pungnap2-dong, Songpa-gu, Seoul, Republic of Korea.
Journal of Digital Imaging
|October 19, 2017
Summary
Deep learning using Convolutional Neural Networks (CNNs) significantly outperformed shallow learning (Support Vector Machines) in classifying interstitial lung disease (ILD) patterns from CT scans, improving accuracy and reducing misclassification rates.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging Analysis
Background:
- Interstitial lung diseases (ILDs) present diverse patterns on high-resolution computed tomography (HRCT).
- Accurate classification of ILD patterns is crucial for diagnosis and treatment.
- Shallow learning models have shown potential but may struggle with complex pattern recognition.
Purpose of the Study:
- To compare the performance of deep learning (CNN) and shallow learning (SVM) for classifying ILD patterns.
- To evaluate the impact of CNN architecture, specifically the number of convolution layers, on classification accuracy.
- To assess the ability of CNNs to differentiate between ambiguous ILD patterns.
Main Methods:
- Utilized 1200 regions of interest (ROIs) from HRCT scans, including normal lung and five ILD patterns (ground-glass opacity, consolidation, reticular opacity, emphysema, honeycombing).
- Developed a six-learnable-layer Convolutional Neural Network (CNN) with four convolution layers and two fully connected layers.
- Compared CNN classification results against a Support Vector Machine (SVM) classifier.
Main Results:
- The CNN classifier achieved significantly higher accuracy (6-9% improvement) compared to the SVM classifier.
- Increasing the number of convolution layers in the CNN improved classification accuracy from 81.27% to 95.12%.
- CNNs demonstrated a reduced misclassification rate, particularly in differentiating between pathologically ambiguous cases like normal vs. emphysema and honeycombing vs. reticular opacity.
Conclusions:
- Deep learning with CNNs offers superior accuracy for classifying ILD patterns from HRCT images compared to SVMs.
- The performance of CNNs in ILD pattern classification is enhanced by increasing the depth of convolution layers.
- CNNs show promise in accurately distinguishing subtle and ambiguous ILD patterns, potentially improving diagnostic capabilities.
