Table 3.

Quantitative performance evaluation of the cell segmentation IDL module on cell tracking challenge test Datasets³⁹.

CTC metrics	Cell tracking challenge testing datasets
	PhC-C2DH-U373	DIC-C2DH-HeLa	Fluo-N2DL-HeLa	Fluo-N2DH-GOWT1	Fluo-N2DH-SIM+
$O{P}_{\mathit{CSB}}$(%)	95 (CALT-US:96.1)	84.3 (CALT-US:92.6)	86.4 (BFR-GE:95.7)	88.1 (KTH-SE:95.2)	84.3 (KIT-GE:90.5)
SEG(%)	91.7 (CALT-US:93.1)	80.1 (CALT-US:88.7)	78.8 (MU-US:92.3)	85.2 (CSU-CN:93.8)	72.2 (DKFZ-GE:83.2)
DET(%)	98.3 (CALT-US:99.0)	88.5 (CALT-US:97.5)	94.1 (KIT-GE:99.4)	91.1 (TUG-AT:98.0)	96.5 (FR-GE:98.1)

This evaluation uses results computed by the challenge organizers³⁹ upon submission of our test set results (where no ground truth was provided). The AttUnet(XAI) performance was assessed using the challenge metrics against the test ground truth masks. The $O{P}_{\mathit{CSB}}$ metric is defined as $0.5\times (DET+SEG)$. Here, $DET=1-\frac{min(AOGM-D,AOGM-D_0)}{AOGM-D_0}$, which represents the normalized Acyclic Oriented Graph Matching (AOGM-D) measure for detection⁴⁰. The SEG measure employs the Jaccard similarity index, expressed as $J(S,R)=\frac{|R\cap S|}{|R\cup S|}$, where R is the set of pixels in a reference object and S is the set of pixels in its matching segmented object. A reference object R and a segmented object S are considered matching if $|R\cap S|>0.5\times |R|$. The top-performing apporach per dataset and per metric are mentioned as follow (approach : score) from CTC website³⁹.