Pathological cerebral atrophy can arise from various neurological and systemic conditions, including depressive disorders, and early identification is clinically important for improving care. However, existing neuroimaging studies primarily rely on magnetic resonance imaging (MRI), which may be less accessible in routine clinical practice. To address this gap, we present an explainable deep learning framework for cerebral atrophy classification from widely available brain computed tomography (CT) images, while incorporating psychological risk factors as auxiliary input features in a multimodal model. The proposed method integrates a convolutional neural network (CNN) to capture structural brain features from CT images with a multilayer perceptron to incorporate psychological risk factors, including depressive symptoms and suicidal ideation. This multimodal design is intended to provide a broader representation of patient characteristics. In addition, gradient-based class activation mapping is employed to provide visual explanations by highlighting anatomically relevant regions that contribute to model predictions.  In the current evaluation setting and at a predefined operating threshold, the image-only model achieved 83.01% accuracy, and the multimodal model achieved 87.25% accuracy. The multimodal approach therefore showed improved performance in this cohort, and attention maps were generally consistent with clinically relevant patterns such as ventricular enlargement and sulcal widening. These findings provide preliminary evidence for the feasibility of CT-based multimodal atrophy assessment, while threshold-independent analyses and broader external validation remain necessary before deployment-oriented conclusions.