Evaluating conditional convolution for magnification-aware histopathological image classification

Abstract

Histopathological breast cancer images exhibit multi-scale variability and class imbalance, challenging automated classification. This study integrates conditionally parameterized convolutions (CondConv) into convolutional neural networks (CNNs) to enhance accuracy and robustness across heterogeneous magnifications. This study proposes CondConv-augmented mobile network version 2 (MobileNetV2) and densely connected convolutional network 121 (DenseNet121), which dynamically adapt convolutional kernels to input characteristics, improving generalization across spatial resolutions. This is the first evaluation of CondConv-based architectures for multi-magnification histopathological classification. This study evaluates these architectures at 40×, 100×, 200×, and 400× using imbalanced and class-rebalanced datasets. Performance was measured via convergence analysis, macro-averaged metrics, and statistical validation with Friedman tests, Nemenyi post-hoc comparisons, and Wilcoxon signed-rank tests. CondConv-enhanced networks matched or surpassed static-kernel variants, with notable gains at 100× and 200× where scale variability peaks. These improvements persisted under class imbalance, demonstrating resilience to distributional skew. Although differences narrowed at 400× due to feature saturation, statistical analyses confirmed significant enhancements at lower magnifications, especially for CondMobileNetV2.