Table 1.
Summary of the TCIA BREAST-DIAGNOSIS dataset characteristics.
Fig 1.
Collection of breast MRI scans, categorized into Benign and Malignant cases.
Fig 2.
AIN Development Process - Sequential four-step development of Adaptive Involution Network for breast MRI tumor classification: introducing SAIL layers, integrating CWAM mechanisms, combining with CNN architecture, and achieving final AIN model.
Fig 3.
SAIL Architecture Overview - Diagram showing the key components of Spatially-Adaptive Involution Layers, including adaptive kernel generation, spatial feature unfolding, involution operations, channel projection, and parallel convolution across multi-scale feature maps.
Fig 4.
CWAM Three-Step Process - Channel-Wise Attention Mechanism showing global context aggregation, channel dependency modeling through bottleneck transformation, and feature recalibration with attention weights.
Fig 5.
Adaptive Residual Block Flow - Circular process showing ARB components: feature extraction through SAIL/convolutions, batch normalization, ReLU activation, dropout regularization, additional convolution, CWAM recalibration, and shortcut connection.
Fig 6.
AIN Sequential Architecture - Gear-based diagram showing AIN processing stages from feature extraction stem through four ARB stages (SAIL-based stages 1-2, convolution-based stages 3-4), global average pooling, to classification head.
Table 2.
Complete list of training hyperparameters used for the proposed AIN model.
Fig 7.
Training and Validation History - Folds 1-3 - Performance plots showing model accuracy and loss progression over training epochs for each fold during cross-validation.
Fig 8.
Training and Validation History - Folds 4-5 and Overall Performance - Performance plots showing model accuracy and loss progression over training epochs for the remaining cross-validation folds plus overall performance summary.
Table 3.
Classification report summary.
Table 4.
Summary of classification metrics across five folds.
Fig 9.
Detailed confusion matrices for breast cancer classification showing individual fold performance (a-e) and aggregated results across all folds (f), demonstrating model consistency and overall classification accuracy.
Fig 10.
Calibration curves illustrating the reliability of predicted probabilities for different models within a five-fold cross-validation framework.
Subfigures (a)–(e) show individual fold results, while (f) presents the overall calibration.
Table 5.
Ablation study on the contribution of SAIL, CWAM (SE), and the hybrid layering strategy.
All models are trained for a maximum of 50 epochs with early stopping (patience = 15).
Fig 11.
Qualitative analysis of model predictions.
Table 6.
Performance comparison of AdaptiveInvolutionNet with state-of-the-art methods for breast cancer classification.