Validated respiratory drug deposition predictions from 2D and 3D medical images with statistical shape models and convolutional neural networks

TL;DR

An image processing approach is proposed that could produce 3D patient respiratory geometries from 2D chest X-rays and 3D CT images and is capable of providing patient-specific deposition measurements for varying treatments to determine which treatment would best satisfy the needs imposed by each patient.

Abstract

For the one billion sufferers of respiratory disease, managing their disease with inhalers crucially influences their quality of life. Generic treatment plans could be improved with the aid of computational models that account for patient-specific features such as breathing pattern, lung pathology and morphology. Therefore, we aim to develop and validate an automated computational framework for patient-specific deposition modelling. To that end, an image processing approach is proposed that could produce 3D patient respiratory geometries from 2D chest X-rays and 3D CT images. We evaluated the airway and lung morphology produced by our image processing framework, and assessed deposition compared to in vivo data. The 2D-to-3D image processing reproduces airway diameter to 9% median error compared to ground truth segmentations, but is sensitive to outliers of up to 33% due to lung outline noise. Predicted regional deposition gave 5% median error compared to in vivo measurements. The proposed framework is capable of providing patient-specific deposition measurements for varying treatments, to determine which treatment would best satisfy the needs imposed by each patient (such as disease and lung/airway morphology). Integration of patient-specific modelling into clinical practice as an additional decision-making tool could optimise treatment plans and lower the burden of respiratory diseases.

Figures (9)

• Figure 1: Schematic overview of image-processing framework. Chest X-ray image(s) are provided as an input (a), where the shape is identified by our SSAM by iteratively adapting lung outline and gray-value based on a database of known lung shapes (b-c). The SSAM reconstruction can then be used to generate a surface mesh of lungs with a full airway tree (c). Alternatively, a 3D CT scan may be provided as an input (d) to our trained CNN architecture (e). The CNN returns a segmented labelmap for lungs and airways (f), which can also be used to generate the full respiratory domain (c).
• Figure 2: Assessment of sensitivity of SSAM metrics to number of modes included in the model. Panels show (a) explained variance based on number of modes, (b) reconstrucion error for shapes in training set, (c) generalisation error for unseen shapes. When the model describes 90% of the population variance (indicated by blue dashed line), the generalisation error in adapting the model to unseen shapes is saturated.
• Figure 3: Lung space volume error from SSAM and U-Net reconstructions of patients from LUNA16 dataset. Inset shows U-Net lung space volume error, where the U-Net results were produced using the pretrained U-Net from hofmanninger2020automatic.
• Figure 4: Example of fitted SSAM compared to projection and lung outline for two largest lung space volume errors. Yellow markers are output landmarks from the SSAM, and black markers represent the X-ray edges detected by the Canny edge map described in Section \ref{['sec:preproc']}. Panels show (a) the largest error, (b) second largest error and (c) the lowest error. In (a) the right lung error was 31% and left lung error was 42.6%. In (b) the right lung error was 15.5% and left lung error was 38.3%. In (c) the right lung error was 1.8% and left lung error was 0.06%.
• Figure 5: Comparison of airway diameter errors predicted by our SSAM compared to the ground truth segmentations. We show the influence of including an additional projection on diameter error in the trachea and main bronchi.