Estimating pulmonary gas-exchange impairment is a clinically relevant yet
challenging task. A standard clinical measure of diffusion capacity of the
lung for carbon monoxide (DLCO) produces a key functional biomarker
reflecting efficiency of alveolar–capillary gas transfer. It is routinely used in
the assessment of diseases such as chronic obstructive pulmonary disease,
interstitial lung disease, and pulmonary vascular disorders. Despite its
widespread use, the various structural sources contributing to DLCO deficit
remain elusive. If it were possible to leverage routinely acquired chest CT
scans to estimate DLCO, this would offer the potential to scrutinize the
relationship between structural imaging and functional impairment,
facilitating a more accurate disease diagnosis.
A critical limitation of purely predictive approaches where global CT metrics
(e.g., % emphysema) are related to DLCO is their lack of specificity: a
single estimated DLCO value measured at the mouth in the lung function
lab can arise from a range of underlying structural drivers of impairment,
depending on their regional distribution across the lungs. In contrast,
models that provide interpretable regional contributions can localize which
lung areas most strongly influence the predicted functional deficit. Such
spatially resolved explanations will likely facilitate hypothesis generation
regarding structure–function relationships, which may help differentiate
disease phenotypes (e.g., emphysema-dominant versus fibrosis-dominant
patterns). More importantly it can support clinical decision-making to guide
therapeutic interventions (e.g., lung volume reduction by resection or
bronchial valves).

A critical limitation of purely predictive approaches is their lack of
interpretability: a single estimated DLCO value provides limited insight into
the underlying structural drivers of impairment. In contrast, models that
provide interpretable regional contributions can localize which lung areas
most strongly influence the predicted functional deficit, thereby enhancing
clinical trust and offering mechanistic insight. Such spatially resolved
explanations may help differentiate disease phenotypes (e.g., emphysemadominant
versus fibrosis-dominant patterns), support clinical decisionmaking,
and facilitate hypothesis generation regarding structure–function
relationships.
Several methodological approaches can be employed to obtain these
explanations. Patch-based or multiple instance learning frameworks
naturally enable aggregation of local image features into a global prediction
while preserving spatial correspondence. Within this setting, attentionbased
pooling mechanisms can assign importance weights to individual
patches, which can be projected back onto the image to generate regionlevel
importance maps. Alternatively, patch-level prediction maps can
visualize local contributions to the global estimate, while gradient-based
techniques such as Grad-CAM provide voxel-level saliency. More
computationally intensive but causally informative methods, such as
occlusion or perturbation analysis, can further quantify the contribution of
specific regions by measuring the change in prediction when they are
masked. Together, these approaches allow the development of models that
not only estimate DLCO but also provide interpretable, spatially resolved
explanations of gas-exchange impairment from CT imaging.

To develop a weakly supervised deep learning model that estimates
diffusion capacity (DLCO) from chest CT scans and clinical metadata, while
providing interpretable regional contributions that link structural lung
abnormalities to functional gas-exchange impairment. We will also
investigate how exploiting the meta data (patient age, BMI, smoking
history, etc…) could provide additional source of supervision in training such
models.

The developments will be performed in close collaboration with the
university hospital UZ Brussel.
The following tasks can be distinguished within the thesis:
• Literature study.
• Implementation: Python
• Validation and testing