Unsupervised machine learning identifies predictive progression markers of IPF

This retrospective study was approved by the Institutional Review Board of the Medical University of Vienna (Ethics Committee number 1463/2017). The local Institutional Review Board waived the informed consent.

Study population

For the study cohort, IPF patients with diagnosis of IPF between December 2011 and October 2014 were retrospectively retrieved from the electronic registers of an Italian referral center (Ospedale Morgagni di Forlì, Italy, n = 76). Inclusion criteria were as follows: (1) availability of at least two consecutive HRCT examinations per patient performed at least 6 months interval; (2) usage of a high-frequency reconstruction kernel (BONEPLUS) with a slice thickness of ≤ 1.25 mm for both examinations. Following these inclusion criteria, a total of 76 patients (f/m: 19/57) were included, as only in these patients follow-up scans with the same reconstruction kernel were available (Fig. 1b). For a sub-cohort of 74 patients, survival data was available. The patient characteristics of the entire study cohort can be seen in Table 1.

Fig. 1figure 1

Overview of the algorithm and dataset. a First, unsupervised learning selects marker candidates, which results in most significant progression markers. b This flowchart represents the selection of enrolled scans

Table 1 Patient characteristics of the entire study cohort

As a replication cohort, we collected a retrospective cohort from a different center and country (n = 18, Vienna General Hospital, Austria). Patients in this dataset were diagnosed with IPF between April 2007 and April 2017. The inclusion criteria were the same as the study referral center, but with a different CT reconstruction kernel (B60f, B70f, B70s, I70f, I80s) since scanners were from a different manufacturer.

For both cohorts, the CT diagnosis was established by two experienced radiologists. The diagnosis of the IPF was made by the multidisciplinary ILDs boards of both institutions.

Imaging data collection and acquisition

The study cohort dataset (Italy) was acquired with 2 CT scanners, a LightSpeed Pro 16, and a BrightSpeed 16 (both GE Healthcare). The CT examinations were performed in supine position in sustained deep inspiration. In case of more than two CT examinations per patient, each pair of consecutive CT scans was included. Therefore, an individual patient could have 1–4 pairs of scans. For the replication cohort, data was acquired with a Siemens Sensation Cardiac 64 scanner in supine position with deep inspiration. Each patient had 2 scans, one at the time of the diagnosis and another one at the last seen examination.

Segmentation and follow-up registration

Lungs were automatically segmented in the CT data in a two-step approach. First, a threshold-based algorithm was used [14], and morphological area opening removed small structures such as small bronchi and vessels. The background was suppressed, and airways were extracted by trachea localization. If the algorithm failed the volume-based assessment criterion, as in cases of substantial high-density areas and lung scarring, a multi-template atlas–based segmentation approach was used to correct the segmentation [15]. This segmentation approach automatically selected an optimal templated lung transformation (VISCERAL Anatomy 3 [16]) using normalized cross correlation criteria and performed a non-linear registration from this selected transformation to a predefined target atlas with Ezyes [17].

After lung mask segmentation, pairs of consecutive CT scans were registered using Advanced Normalization Tools (ANTs) [18] to establish correspondence of positions in the lung imaging data for subsequent examinations of the same patient. Registration of corresponding positions was necessary to track the change of lung pattern classes over time and measure local transitions between pattern classes during radiological disease progression.

To reduce computational complexity, instead of processing each individual voxel of the slices, we over-segmented the lung mask into small parcels of the size of 5 mm × 5 mm × 5 mm — so-called supervoxels — using MonoSLIC [19]. Those supervoxels were extracted through k-means clustering performed on the monogenic phase detecting the locally dominant structure of the CT voxel regardless of the contrast and brightness of the image, resulting in a total of 1,578,788 supervoxels covering the entire lung cohort (Fig. 1a).

Extraction of CT features and radiological disease progression marker candidates

We identified distinct lung appearance patterns occurring across the entire study population by unsupervised machine learning on all imaging data using a bag of visual words approach [20]. Computed tomography imaging data was received in the form of DICOM files, with gray values representing Hounsfield units. The gray value range was transformed to 0 to 255 before extracting image features. We calculated various statistical properties of the orientation-independent gray-level co-occurrence matrices for each supervoxel, resulting in so-called Haralick textural features [21], a 65-component vector per supervoxel. To reduce this high dimensionality, we used principal component analysis (PCA) retaining 95% of the overall variance resulting in a 9 dimensional feature vector per supervoxel (Fig. 1a).

K-means clustering in this 9 dimensional feature space assigned each supervoxel to one lung appearance pattern corresponding to a cluster. The optimal number of clusters was determined by repeatability testing, using Jaccard score [22] resulting in k = 20 clusters as the optimal choice. Finally, each lung was represented by the volume fraction covered by each of the 20 appearance patterns, resulting in a vector of 20 components. This pattern signature captures the overall texture composition of the lung.

Identifying marker patterns of radiological disease progression and pathways of local tissue transition

IPF is associated with pulmonary fibrosis, a type of terminal pathological change in the lung, caused by chronic repetitive alveolar injury and results in excessive synthesis of extracellular matrix and replacement of normal parenchyma. While some types of pulmonary fibrosis are reversible, IPF exhibits progressive and irreversible development [23, 24].

To identify pattern signature components associated with radiological disease progression, we analyzed available pairs of subsequent CT scans of the same patient (scan A and scan B) with known acquisition dates. We trained a random forest (RF) [25] classification model with 500 trees to predict the correct temporal sequence of two scans (A, B) based on the difference of their pattern signatures (prediction result: A acquired after B or B acquired after A, ground truth during training based on the acquisition dates in the DICOM header). We used Gini importance to rank features regarding their predictive power for correct sorting. Since IPF is irreversible, the lung scarring captured in CT scans either remains the same or worsens. Thus, we hypothesize that features enabling correct temporal sorting capture radiological disease progression.

The ground truth for the correct temporal sequence was read from the CT DICOM-header. We used RF Gini importance to score the contribution of pattern signature components to the correct sorting of scans, and hypothesize that components — each associated with a lung tissue type with specific appearance — with high score are strongly associated with radiological disease progression.

Predicting outcome based on the dynamics of pattern signatures

We used only the top 4 scored components determined previously together with their change between a pair of follow-up scans to form the radiological disease progression signature. We clustered patients into two groups using k-means clustering of their progression signatures. For each patient cluster, we assessed survival in a Kaplan-Meier analysis.

Exploratory analysis of pattern transition pathways

We analyzed if the transition of lung tissue from one to a different pattern follows one or more specific sequences during the course of the disease. We determined the image signature component at each lung position in one scan, and the component at the corresponding position in the subsequent scan for all 1,578,788 supervoxels and all scan pairs in the study. This yielded a transition probability network. It captures how likely it is to transition from one of the 20 tissue patterns to another during radiological disease progression.

Evaluation

To validate the radiological disease progression model, we tested if the machine learning model could correctly determine the temporal sequence of pairs of subsequent CT examinations based solely on the image signatures extracted from each of the two CT volumes, resulting in the “sorting accuracy” of the RF model.

To evaluate if inaccuracies stem from a lack of visible radiological disease progression, or algorithmic limitations, we compared the sorting accuracy of the RF model with the sorting accuracy of two experts with 17 years (expert 1) and 15 years (expert 2) of experience in thoracic radiology. The radiologists were shown pairs of follow-up scans in random order blinded for examination dates. Algorithm sorting accuracy was evaluated in leave-one-patient-out cross-validation, by training the machine learning model, and identifying radiological disease progression markers on 76 patients of the dataset, and automatically sorting the remaining pair of scans from a patient of the dataset with the trained RF model.

To assess the stability of radiological disease progression marker patterns, we randomly picked 20 subsets of 95% (n = 72) of the patient’s scan pairs in each run, to train the machine learning model, and tested if the ranking of the top informative radiological disease progression markers remained the same. The top-ranked prototypes were assessed and evaluated as image patches (250 × 250 pixels) by an expert for their content.

To evaluate if the progression signatures predict outcome, we assessed the hazard ratio (HR) between the two patient clusters identified based on the progression signatures. In the study cohort, for 74 patients, survival data was available, and analysis was performed on those patients. In a replication experiment, we processed the external validation data (n = 18) using the same 4 components of the progression signature, and assigned each new patient to one of the two existing patient clusters identified in the study cohort. We evaluated replicability by Kaplan-Meier analysis, analogously to the study cohort.

留言 (0)

沒有登入
gif