TEMPORAL STRIDE AS AN INDUCTIVE BIAS FOR SE- MANTIC LEARNING IN MEDICAL VIDEOS

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: tiny / short paper (2-4 pages excluding references; extended abstract format)

Keywords: Self-supervised learning, Video representation learning, Temporal stride, Inductive bias, Masked autoencoders, Medical imaging, Colonoscopy, Anomaly detection.

Abstract:

Self-supervised video representation learning has become a standard paradigm in
medical imaging, yet the semantic implications of the temporal sampling stride
remain unexplored. In this work, we investigate whether the temporal spacing
between training frames acts as an inductive bias that shapes the learned repre-
sentations. We conduct a controlled study using VideoMAE v2 on full-length
colonoscopy videos, comparing encoders pre-trained with dense sampling (stride
1) versus sparse sampling (stride 30). Our analysis reveals a fundamental trade-
off: dense sampling yields representations dominated by low-level appearance
and motion features, resulting in a polyp detection F1-score of 60.3% and a near-
random anomaly detection AUC of 0.449 when using a Euclidean distance. Con-
versely, sparse sampling forces the encoder to capture anatomical semantics, im-
proving F1-score by 16.5 percentage points and achieving near-perfect anomaly
detection under Mahalanobis scoring (AUC = 0.998).

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Ahmad_Taha3

Format: Maybe: the presenting author will attend in person, contingent on other factors that still need to be determined (e.g., visa, funding).

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 58