Table 2 Results of the proposed approach for BC classification alongside the different ablation studies.

From: Embedding-driven dual-branch approach for accurate breast tumor cellularity classification

Approach

Accuracy (%)

Sensitivity (%)

Specificity (%)

Precision (%)

F1 (%)

Average (%)

Proposal (full model)

\(97.86 \pm 0.0013\)

\(97.86 \pm 0.0013\)

\(99.29 \pm 0.00043\)

\(97.86 \pm 0.00131\)

\(97.86 \pm 0.0013\)

98.15

 

[97.85–97.87]

[97.85–97.87]

[99.29–99.30]

[97.85–97.87]

[97.85–97.87]

 

No data augmentation

\(89.37 \pm 0.00523\)

\(89.37 \pm 0.00523\)

\(96.46 \pm 0.00174\)

\(89.37 \pm 0.00531\)

\(89.37 \pm 0.00531\)

90.79

 

[89.36–89.38]

[89.36–89.38]

[96.45–96.47]

[89.36–89.38]

[89.36–89.38]

 

No embedding branch

\(25.00 \pm 0\)

\(25.00 \pm 0\)

\(75.00 \pm 0\)

N/A

N/A

41.67

 

[25.00–25.00]

[25.00–25.00]

[75.00–75.00]

–

–

 

No vision branch

\(95.75 \pm 0.00248\)

\(95.75 \pm 0.00248\)

\(98.58 \pm 0.00083\)

\(95.79 \pm 0.00249\)

\(95.77 \pm 0.00248\)

96.33

 

[95.74–95.76]

[95.74–95.76]

[98.57–98.59]

[95.78–95.80]

[95.76–95.78]

 
  1. Values are reported as mean ± standard error across 10 independent trials. 95% confidence intervals (CIs) are provided in square brackets. The “No Embedding Branch” condition shows deterministic performance due to complete model failure (equivalent to random guessing). Precision and F1 are marked N/A when true positives = 0