Domain knowledge acquired from pilot studies is important for medical diagnosis. This paper leverages the population-level domain knowledge based on the D-optimal design criterion to judiciously select CT slices that are meaningful for the disease diagnosis task. As an illustrative example, the diagnosis of idiopathic pulmonary fibrosis (IPF) among interstitial lung disease (ILD) patients is used for this work. IPF diagnosis is complicated and is subject to inter-observer variability. We aim to construct a time/memory-efficient IPF diagnosis model using high resolution computed tomography (HRCT) with domain knowledge-assisted data dimension reduction methods. Four two-dimensional convolutional neural network (2D-CNN) architectures (MobileNet, VGG16, ResNet, and DenseNet) are implemented for an automatic diagnosis of IPF among ILD patients. Axial lung CT images are acquired from five multi-center clinical trials, which sum up to 330 IPF patients and 650 non-IPF ILD patients. Model performance is evaluated using five-fold cross-validation. Depending on the model setup, MobileNet achieved satisfactory results with overall sensitivity, specificity, and accuracy greater than 90%. Further evaluation of independent datasets is underway. Based on our knowledge, this is the first work that (1) uses population-level domain knowledge with optimal design criterion in selecting CT slices and (2) focuses on patient-level IPF diagnosis.

Lack of classifier robustness is a barrier to widespread adoption of computer-aided diagnosis systems for computed tomography (CT). We propose a novel Robustness-Driven Feature Selection (RDFS) algorithm that preferentially selects features robust to variations in CT technical factors. We evaluated RDFS in CT classification of fibrotic interstitial lung disease using 3D texture features. CTs were collected for 99 adult subjects separated into three datasets: training, multi-reconstruction, testing. Two thoracic radiologists provided cubic volumes of interest corresponding to six classes: pulmonary fibrosis, ground-glass opacity, honeycombing, normal lung parenchyma, airway, vessel. The multi-reconstruction dataset consisted of CT raw sinogram data reconstructed by systematically varying slice thickness, reconstruction kernel, and tube current (using a synthetic reduced-tube-current algorithm). Two support vector machine classifiers were created, one using RDFS ("with-RDFS") and one not ("without-RDFS"). Classifier robustness was compared on the multi-reconstruction dataset, using Cohen's kappa to assess classification agreement against a reference reconstruction. Classifier performance was compared on the testing dataset using the extended g-mean (EGM) measure. With-RDFS exhibited superior robustness (kappa 0.899-0.989) compared to without-RDFS (kappa 0.827-0.968). Both classifiers demonstrated similar performance on the testing dataset (EGM 0.778 for with-RDFS; 0.785 for without-RDFS), indicating that RDFS does not compromise classifier performance when discarding nonrobust features. RDFS is highly effective at improving classifier robustness against slice thickness, reconstruction kernel, and tube current without sacrificing performance, a result that has implications for multicenter clinical trials that rely on accurate and reproducible quantitative analysis of CT images collected under varied conditions across multiple sites, scanners, and timepoints.