Background
Dysphagia, or difficulty swallowing, affects about one in eight Dutch adults, rising to one in four of those over 76. It can lead to aspiration pneumonia, malnutrition and loss of independence. The videofluoroscopic swallow study (VFSS), an X-ray video of the swallow, is the gold standard for diagnosing it.
In routine care, VFSS is interpreted by eye, and visual interpretation is unreliable: experienced clinicians classified only 61.5% of swallowing parameters correctly against objective measurement. Objective measurements exist but take too long to perform manually, so they are rarely used. Deep learning has been applied to VFSS, but only for single tasks on single-centre data, and no end-to-end tool exists yet.
We are starting the development of a foundation model for VFSS: one shared backbone, initialized from a pretrained medical imaging foundation model, with separate task heads that extract all relevant swallowing parameters in one pass.
Aim
In this assignment, we aim to answer find out which deep-learning head works best for one of the most clinically important task families: the timing of swallowing events, such as laryngeal vestibule closure and upper oesophageal sphincter opening. The assignment has three phases: a literature review of the deep learning models validated for VFSS and the tasks they address; annotation of swallowing events in our VFSS dataset, together with a speech and language therapist; and the implementation and evaluation of deep-learning architectures for the detection of quantitative swallowing metrics.
