Ovarian tumors pose a critical threat to women’s health due to their potential malignancy and impact on fertility. Early and accurate classification of ovarian tumors is vital for timely treatment, yet most existing deep learning models for medical imaging remain limited by static convolutional structures and high computational demands. This research addresses these limitations by developing an adaptive framework capable of dynamically optimizing convolutional operations based on input characteristics.
We propose a novel Dynamic Dimensional Convolutional Network that introduces a dual-module mechanism: a dimension-selection module and a filter-selection module. These components empower the network to automatically adjust convolutional dimensions and filter sizes during training and inference. The model was trained and evaluated on a 1200-sample ovarian ultrasound dataset using TensorFlow and Keras, and its performance was compared against ResNet, EfficientNetB4, and EfficientNetV2 baselines.
Quantitative analysis shows the the proposed model accomplished an average classification correctness of 97.6%, outdoing existing Convolutional neural network-based models by 5–6% while diminishing computational cost by approximately 12% in floating-point operations. Qualitative analysis confirmed enhanced feature discrimination and toughness under different imaging situations.
This analysis introduces a dynamic dimensional convolutional approach that unifies Unidimensional, Bidimensional, and Tridimensional kernel operations within a single lightweight model. Major contributions consist of: (1) an adaptive mechanism for dimension and filter selection, (2) better accuracy with reduced computation, and (3) a scalable framework suited for real-time medical imaging tasks.
The proposed DyDimNet framework delivers a dynamic dimensional convolutional approach that significantly improves both accuracy and performance in ovarian tumour classification. The DyDimNet automatically determines the ideal parameter sizes and configurations. This makes it compatible with many types of images and allows for detailed spatial information extraction. This flexibility increases the accuracy of the tumor detection, with lower computing load. The experimental results show that DyDimNet is more accurate than the traditional deep learning models and easy enough to apply in the clinical setting. This method balances the need for accurate diagnosis with real-world computer limits, making it a promising choice for bringing smart image analysis into healthcare.