LiteRenalNet: A lightweight dual-branch network with channel–spatial co-attention for interpretable kidney CT classification

Abstract Automated classification of renal computed tomography (CT) can accelerate triage and extend reliable screening to settings where specialist expertise is scarce, yet the most accurate deep models rely on heavy general-purpose backbones, exploit only a single visual scale, and offer little insight into the basis of their predictions. We propose LiteRenalNet , a lightweight and interpretable dual-branch convolutional neural network that classifies an axial renal CT slice as normal, cyst, stone, or tumor. The network couples a shallow high-resolution detail branch, which preserves the high-frequency cues of small lesions such as calculi and thin cyst walls, with a deep low-resolution semantic branch, which encodes the global morphology of larger masses, and combines them through a bilateral co-attention fusion module. At its core is a Channel–Spatial Co-Attention (CSCA) module that, unlike conventional sequential attention, estimates channel and spatial attention along parallel paths and fuses them into a single joint three-dimensional attention map, with a zero-initialized gating that stabilizes early training. Built almost entirely from depthwise separable convolutions and inverted residual units, LiteRenalNet contains only 1.27 M parameters yet attains $$99.25\\%$$ 99.25 % test accuracy, a macro-F1 of 0.9913, and a macro-AUC of 0.9993, misclassifying just 14 of 1, 867 slices, with clinically hazardous confusions limited to a single low-confidence tumor-as-normal case. A region-level Shapley value analysis further confirms that the model’s decisions are concentrated on the retroperitoneal renal region rather than on background artefacts, demonstrating that its high accuracy reflects clinically meaningful and trustworthy reasoning.