Kidney Tumor Detection in CT scans Combining 3D CNN and Transformer Networks

This paper introduces a hybrid deep learning architecture for kidney tumor detection in CT scans, integrating 3D convolutional operations with transformer-based sequence modeling. Our volumetric processing framework combines the local feature extraction capabilities of 3D CNNs with the global contextual understanding of transformers, specifically designed for multi-slice CT analysis. The system processes 16-slice CT volumes through a ConvNeXt-inspired 3D backbone that maintains spatial relationships across imaging planes, followed by a transformer encoder that captures long-range dependencies between slices through learned attention patterns.

The architecture employs several key innovations: (1) a depth-aware patch embedding layer that preserves inter-slice relationships when transitioning from 3D convolutions to 2D attention maps, (2) adaptive positional encodings that account for variable slice spacing in medical CT acquisitions, and (3) a multi-scale feature fusion module that combines hierarchical representations from both pathways. We demonstrate the system's clinical utility through interpretable attention visualizations that highlight diagnostically relevant regions across slice sequences.