Table 16 Ablation study results for MIAS dataset.
Model Variation | Optimizer | Accuracy (%) | Precision (%) | Recall (%) | F1-Score (%) | Comments |
|---|---|---|---|---|---|---|
Full Model (CNN + EfficientNet-B0 + Bi-LSTM) | Adam | 99.2 | 98.7 | 99.5 | 99.1 | Best performance across MIAS dataset, combining feature extraction and temporal modeling. |
CNN + EfficientNet-B0 (Without Bi-LSTM) | Adam | 97.5 | 96.8 | 98.0 | 97.4 | EfficientNet-B0 provides intense feature extraction but lacks temporal analysis. |
CNN + Bi-LSTM (Without EfficientNet-B0) | Adam | 95.8 | 94.5 | 96.2 | 95.3 | Temporal modeling with Bi-LSTM alone still provides exemplary accuracy. |
CNN Only (Without Bi-LSTM or EfficientNet-B0) | Adam | 91.6 | 89.4 | 92.1 | 90.7 | Baseline model without advanced feature extraction or temporal data processing. |
Full Model (CNN + EfficientNet-B0 + Bi-LSTM) | RMSProp | 98.9 | 98.0 | 99.1 | 98.6 | RMSProp still performs well, but Adam is superior. |
Full Model (CNN + EfficientNet-B0 + Bi-LSTM) | SGD | 97.2 | 96.0 | 97.5 | 96.7 | SGD shows lower performance due to slower convergence compared to Adam and RMSProp. |
