ROBUST BEHAVIORAL PHENOTYPING UNDER OCCLUSION IN STANDARD IVC RACKS: ADAPTING CROSS-MODALITY DISTILLATION WITH LABEL CORRELATION

Aims: Standard Individually Ventilated Cages (IVC) create visual occlusion that limits the reliability of 24/7 home-cage monitoring. To address this without expensive retrofitting, we developed a framework adapting the cross-modality distillation strategy described by Xu et al. (2021) to compensate for single-view blind spots.
Methods: We implemented a strategic hardware plan tailored to IVC constraints: accessible cages on the rack periphery are designated as dual-view "special care" nodes. These serve as "Teachers," capturing synchronized multi-angle features to train models to recognize occluded behaviors. This learned spatial knowledge is then distilled to restricted single-view "Student" nodes typical of mass deployment. To further mitigate visual ambiguity, we integrated Modelling Label Correlation, utilizing temporal context to assist judgment during occlusion. We are validating this system via an ongoing comparative study using Taiwan’s endemic species, Taiwania cryptomerioides, as environmental enrichment. This design aims to establish local husbandry standards while testing the system's sensitivity to novel environmental stimuli.
Results: Current phase validation indicates that the distillation architecture effectively transfers spatial awareness from Teacher setups to standard Student models. The inclusion of Label Correlation improves inference stability during occlusion events. Initial data suggests the system can detect subtle activity shifts in the Taiwania-enrichment group. We are presently analyzing the correlation between these digital biomarkers and physiological safety outcomes to distinguish benign adaptation from stress.
Conclusions: Preliminary findings support the feasibility of this framework. By combining practical hardware planning with context-aware algorithms, our approach enables robust, non-invasive phenotyping in standard facilities, supporting the 3Rs refinement principle.