AI-Driven Model for Automatic Emphysema Detection in Low-Dose Computed Tomography Using Disease-Specific Augmentation

Medical ethics committee approval was obtained prior to the study. The population for this retrospective study was chosen from two different cohorts. The first was a general population study in the Netherlands (Imaging in Lifelines or ImaLife) designed to find early imaging biomarkers for the “big-three” thoracic diseases, COPD, coronary artery disease, and lung cancer. The second was the National Lung Cancer Screening Trial (NLST) which was carried in the USA. It is one of the biggest collections of lung cancer screening LDCT data and contains case-specific datasets with annotations. The details regarding the eligibility criteria for the participants of ImaLife and NLST have been previously described [25, 26].

The CT acquisition protocol for the ImaLife study participants included a low-dose chest CT examination with a third-generation dual-source CT scanner (SOMATOM Force, Siemens). All the scans were reconstructed in the axial plane with filtered back projection using a soft kernel (Br40) with a single slice thickness of 1 mm and 0.7-mm increments [25]. The NLST subcohort contained LDCT scans from various hospitals where multi-detector scanners from GE medical systems, Siemens, Toshiba, and Philips with varying slice-thicknesses were used to produce the scans [27]. A brief overview of CT acquisition and reconstruction protocols used in the subcohorts is shown in Table 1. All the scans used in this study were acquired during end-inspiration breath-hold and without the administration of any contrast media.

Each scan from ImaLife was visually scored by one of three trained medical professionals (two radiologists with more than 10 years of experience and one trained technical physician). The three readers followed a standard annotation protocol based on the Fleischner criteria [28]. For this project, the visual scoring was consolidated to dichotomous emphysema diagnosis per scan. The scans with no emphysema annotation were considered normal, and the severity categories trace, mild, moderate, confluent, and advanced destructive emphysema were considered abnormal. For NLST subcohort, radiologists from various hospitals participating in the NLST screening trial had annotated the scans with a yes or no diagnosis for emphysema, and this information was directly used for the external validation [27].

To measure the quantitative characteristics of the subcohorts and evaluate the potential of DL algorithm trained on visual scoring over traditional quantitative analysis, we performed quantitative CT measurements of emphysema using the routinely used automatic densitometry tool (version 4.4.13, Aquarius iNtuition, TeraRecon). The lung region was semi-automatically segmented, and emphysema was quantified as the percentage of all lung area (voxels) having attenuation lower than −950HU (%LAA). Participants with %LAA were divided into two subgroups for analysis, with %LAA ≤ 5% categorized as non-emphysema and %LAA > 5% categorized as emphysema [29].

MinIP is a volume rendering technique where the voxels with lowest attenuation in an image slice are projected to form a bidimensional slab. The slab thickness can be varied based on the number of slices used. First, the acquired scans were re-sampled and normalized to compensate for kernel differences [30]. Then the voxels with lowest Hounsfield units on each slice of a patient scan were projected to form varying slab-thickness. In this study, we generated minIP slabs with thicknesses ranging from 1 up to 11 mm in 2-mm increments for the comparative evaluation. In routine evaluation for emphysema, radiologists use 5 to 10 mm minIP slabs on every slices to carefully check for low attenuation areas [21, 31]. Our study evaluated a wider range to assess the effectiveness of the DL algorithm for each slab thickness separately. Illustration of emphysema case with different MinIP settings with slab thicknesses varying between 1 and 100 mm is shown in Fig. A of Online Resource 1. The algorithm for minIP is implemented in Python and will be made available upon request.

For the development of the DL model, the subcohorts were divided into three datasets. The first was a dataset containing non-emphysema scans for adversarial auto-encoder training, the second was a class-balanced dataset for internal validation, and the third was a class-imbalanced (higher number of non-emphysema scans) dataset for external validation. A total of 160 (80%) out of 200 non-emphysema scans from the ImaLife subcohort were used as the training dataset. The internal validation dataset contained the remaining 40 (20%) normal scans and 40 emphysema scans from the ImaLife subcohort. The complete NLST subcohort was used as the external validation dataset. The consort diagram illustrating the data streams for training and internal and external validation is shown in Fig. 1. It is important to note that the above procedure was followed separately for each minIP slab thickness.

To build a prototype for automatic classification of emphysema in an annotation-less, class-imbalanced environment, we used an unsupervised anomaly detection scheme. The model takes minIPs of certain slab thickness as inputs with a dimension of (512 × 512) × n sampled evenly over the height of the lungs, where n represents the number of axial slices for each participant. The predictions are then concatenated in a sequence of axial slices for the final participant score. The model performs classification on the participant level, so that for a participant to be classified as negative, all slices considered from each participant scan had to be classified as non-emphysema. Otherwise, the participant was classified as positive. The complete DL algorithm was implemented on PyTorch (version 5.1, Python 3.7.1, CUDA 10.1) and executed on a NVIDIA Titan XP GPU with 16 GB memory.

The classification network was built using adversarial auto-encoders, which can be divided into a generator block and a discriminator block (Fig. 2). The first part of the generator block consists of an encoder with fully convolutional layers capable of encoding the high-dimensional image-representation into a low-dimensional latent-representation. The second part of the generator architecture is a decoder that can decode the low-dimensional latent-representation back into a high-dimensional image-representation (X’) [32]. The generator architecture includes sixteen fully convolutional downsampling and upsampling layers with a kernel size of 4 × 4 and stride 2, where each layer is followed by batch normalization and Leaky ReLU activation functions. The details of architectural components and the loss functions have been defined earlier [33]. The encoder’s feature maps are forwarded to the decoder via skip connections with dropout regularization to translate intrinsic image information between blocks without overfitting. This feature maps from the decoder are fed to the discriminator block and the prediction outputs are obtained as anomaly score for each participant scan. The discriminator block has a similar architecture as the generator’s encoder, consisting of eight fully convolutional downsampling layers with a kernel size of 4 × 4 and stride 2, where each layer is followed by batch normalization and Leaky ReLU activation function. We used the discriminative image features from the last convolutional layer of the discriminator as a priori to decide the candidate/anomaly regions in detection maps. Additionally, to train the model with a limited dataset and avoid overfitting, we incorporated discriminator heuristics augmentation [34].

During training, the model was fed with axial slices of given minIP slab thickness in which emphysema was not present so that the generator and the discriminator combination could learn to map the intrinsic properties of non-emphysema lungs. The training was based on minimizing the three loss functions, namely, contextual loss based on L1 distance (or generator’s loss), latent loss (L2 distance measure), and adversarial loss (or discriminator’s loss). The definitions of these three loss functions have been previously published [32].

The trained network in the inference phase classifies the abnormal emphysema scans as anomalies and calculates a prediction score based on generator loss and latent loss functions. The prediction score is the prediction probability provided by the discriminator as a continuous variable ranging from 0 to 1, with higher scores corresponding to emphysema detected.

Along with the prediction scores, the modified network was designed to provide binary masks or residual images for every test image. This was done by subtracting the input image (X) and the generated image (X’) from the generator block [35]. Afterwards, these residual images were post-processed using a lung lobe segmentation algorithm to automatically remove anything apart from the candidate regions generating detection maps. We used an available pre-trained deep learning algorithm trained on the Lung Tissue Research Consortium dataset to perform the post-processing task [36]. The detection maps were then superimposed on the input image to serve as a quality check.

Our model evaluation was performed in three different stages:

The ImaLife subcohort used for training and internal validation consisted of 240 participants (male = 116 [48.4%] and female = 124 [51.6%]); the mean age ± SD at enrollment was 57 ± 6 years. Visual scoring by radiologists indicated 40 (17%) individuals with emphysema in the subcohort. The quantitative CT measurement of emphysema (%LAA −950) in the ImaLife subcohort showed a median of 4% with an interquartile range of 1.8 to 7.8%. Out of 240 individuals, 136 (56.6%) had %LAA ≤ 5%, and 104 (43.4%) had %LAA > 5%.

The NLST subcohort contained 125 (male = 79 [63.2%] and female = 46 [36.8%]) patients with mean age ± SD of 64 ± 5 years at enrollment. The emphysema annotation of the NLST subcohort indicated that 33% individuals had emphysema. The quantitative CT measurement subcohort had median of 15.1% (interquartile range, 5.3 to 28.3%) of emphysema. Out of 125 patients, 25 (20%) had %LAA ≤ 5% and 100 (80%) had %LAA > 5%.

The population characteristics of the subcohorts are shown in Table 2.

The quality check of the detection maps from the DL model by the radiologists revealed that the prediction regions in detection maps were accurate in highlighting emphysema regions in 97% (34 out of 35) and 91% (30 out of 33) of the predicted cases in internal and external-validation datasets, respectively. An example of a detection map is illustrated in Fig. 5.