Validation of the DeepLabCut-based automated method for the novel object recognition test in rats

The novel object recognition (NOR) test is widely used to assess object recognition memory in rodents, but manual scoring is labour-intensive and susceptible to interobserver variability.

We developed an open-source tool combining DeepLabCut (DLC) with explicit numerical criteria. DLC estimated nose, head, and object coordinates in Sprague-Dawley rats. A Python algorithm classified exploration using grid-search-optimised distance, angle, and likelihood thresholds. Low-likelihood frames, including those involving object occlusion during climbing, were excluded.

On an independent dataset of 18,000 frames, sensitivity and positive predictive value were 97% and 83% for the novel object and 97% and 88% for the familiar object, respectively. Across 24 NOR sessions, automated measurements showed high agreement with the mean scores of two independent blinded observers for novel object exploration time (r = 0.87; ICC(2,1) = 0.86), familiar object exploration time (r = 0.96; ICC(2,1) = 0.95), and the novelty discrimination index (NDI) (r = 0.95; ICC(2,1) = 0.95). Bland-Altman analysis showed no evidence of fixed or proportional bias; the 95% limits of agreement for NDI were -0.09-0.09.

The method uses explicitly reported numerical thresholds that can be independently verified and recalibrated. Agreement between automated and manual scoring was assessed using ICC, and systematic bias using Bland-Altman analysis. Proprietary analysis software was not required.

The DLC-based method showed strong agreement with manual scoring under the conditions tested and provides a transparent, accessible approach to automated NOR analysis.