E1621. Automated Classification of Interstitial Lung Disease on High-Resolution Chest CT
Authors
Niharika Shahi;
McMaster University
Abdullah Alabousi;
McMaster University; St. Joseph's Healthcare Hamilton
Ehsan Haider;
McMaster University; St. Joseph's Healthcare Hamilton
Victoria Tan;
McMaster University; St. Joseph's Healthcare Hamilton
Amna Al-Arnawoot;
McMaster University; St. Joseph's Healthcare Hamilton
Oleg Mironov;
McMaster University; St. Joseph's Healthcare Hamilton
Objective:
The aim of our study was to develop an automated model to detect and classify cases of interstitial lung disease (ILD) such as usual interstitial pneumonia (UIP) and nonspecific interstitial pneumonia (NSIP) on high-resolution chest CT (HRCT).
Materials and Methods:
A total of 183 HRCT images of 7 types of ILD were used along with a set of 43 normal images. Three representative lung window axial images from each scan were used, including the upper, mid, and basilar lungs. The images were split into training and validation data sets within a ratio of 80:20. A pretrained EfficientNetV2S convolutional neural network model based on weights obtained for ImageNet was loaded from the TensorFlow environment. The top layer was removed. The network was connected to a distributed layer to apply the pretrained section to each input image. The results were flattened, collated, and fed into a single densely connected layer. The pretrained model expects a 224 x 224 x 3 channel input. The images were down-sampled to 224 x 224 using simple averaging. The same grayscale input was passed to all 3 channels. The model was trained using sparse categorical cross entropy as the loss function, and the result was compared to the validation data set on standard machine learning metrics including accuracy. An initial callback was utilized with a patience of 1 epoch and a minimum delta in accuracy of 0.001. The final layers of the pretrained model were then unlocked, and additional fine tuning was performed with a learning rate of 0.0001 and a patience of 3 epochs.
Results:
After 9 epochs, the validation accuracy on the full data set was 84% for the detection and classification of ILD on HRCT into one of seven categories. When the data were restricted to UIP, NSIP, other, and normal, the validation accuracy increased to 96%.
Conclusion:
A convolutional neural network using transfer learning can achieve high accuracy in the detection and classification of ILD on HRCT. Accuracy is higher when restricted to categories with a larger number of cases.
AI-based detection and characterization of ILD on HRCT is accurate and can have significant implications for aiding radiologist-drive interpretation.
