Computational pathology with dynamic convolutional and adaptive kernels.

Data processing and learning have become essential to the advancement of medicine, with pathology and lab medicine being no exception. Integrating scientific research with clinical informatics into clinical practice facilitates novel methodologies for patient care. Computational pathology is a burgeoning subspecialty in pathology that promises a better-integrated solution to histopathological images and clinical informatics. Deep-learning methods in computational pathology have demonstrated considerable advances in automated histopathological image analysis. However, convolutional neural networks (CNNs) face fundamental limitations when dealing with the significant morphological heterogeneity present in disease tissues. Conventional CNNs use fixed convolutional kernels, which restrict their effectiveness in adaptively extracting features from histopathological images that exhibit diverse pathological patterns, staining intensities, and tissue architecture. To address this substantial limitation, we present an optimized variant of Omni-Dimensional Dynamic Convolution (ODConv) networks for distinguishing diseased tissue from healthy tissue. Compared with prior dynamic convolution methods that attend to a single kernel dimension, ODConv applies multi-dimensional attention across spatial positions, input channels, output channels, and kernel candidates, enabling more flexible and adaptive feature extraction. We evaluated our approach on wheat-germ agglutinin-stained and hematoxylin and eosin-stained skeletal muscle images from multiple disease models, including G93A*SOD1 transgenic mice (amyotrophic lateral sclerosis) and Akita mice (Type I diabetes). ODConv, trained entirely from scratch without ImageNet pretraining, achieved competitive classification performance relative to seven fine-tuned pretrained architectures across both staining modalities, demonstrating the effectiveness of omni-dimensional dynamic kernels in learning discriminative morphological representations directly from domain data. The study reports strong statistical agreement metrics, proving effective class balance handling and stable decision boundaries. These findings confirm ODConv as a strong computational pathology framework that advances automated diagnosis of neurodegenerative and metabolic skeletal muscle disorders.