Cell-TRACTR: A transformer-based model for end-to-end segmentation and tracking of cells
Fig 3
Cell-HOTA metric incorporates cell division accuracy.
(A) Kymograph of bacteria growing in the mother machine. Colored triangles indicate potential times at which a cell division could have occurred, where different colors correspond to different cells. Multiple triangles of the same color indicate that it can be subjective to determine where division occurs. (B) The intersection over union (IOU) between a tracker cell and ground truth cell must exceed the similarity threshold α to be considered a match. HOTA is evaluated across various values of α, with a higher α imposing stricter criteria for matching. (C) Schematic showing the criteria required to determine if a division qualifies as a true positive division (TPD). A tracker cell and ground truth cell are shown dividing at time t1. First, the set of daughter cells that divided at time t1 needs to match. Second, the parent cells of those daughter cells at time t0 needs to match. Third, the average IOU between the set of daughter cells needs to exceed the similarity threshold α. Red lines indicate matching cells. The blue dotted line indicates the set of cells used to calculate the IOU to determine if the division is a TPD. (D) An early or late division occurs when a cell division is predicted one frame too early or late. Black lines indicate temporal associations and brown lines are used to highlight division events. (E) DivA rewards division events within a frame of the ground truth. Flexible divisions may be disabled so that divisions events are only counted as correct if they occur in the same frame as the ground truth. No Flex indicates that DivA has disabled flexible divisions.