This study explores the use of Convolutional Neural Networks (CNNs) to analyze House-Tree-Person (HTP) drawings for the classification of depression severity, addressing the subjectivity and limitations of traditional psychological assessment methods.

A dataset of 1,020 HTP drawings from adults aged 25–30 was collected, consisting of 432 healthy controls, 336 patients with moderate depression, and 252 patients with severe depression. The drawings were labeled based on the Hamilton Depression Scale (HAMD). A CNN model was trained and optimized using cross-validation to extract and classify depression-related visual features. The model's performance was evaluated using accuracy, recall, F1-score, and area under the ROC curve (AUC).

The CNN model demonstrated a classification accuracy of 89% for distinguishing normal and depressed individuals, with an AUC of 0.96. In differentiating moderate from severe depression, the model achieved an AUC of 1.00, indicating near-perfect classification. The extracted features, such as line clarity and detail richness, correlated with depression severity, confirming their diagnostic relevance.

The study validates CNN-based image analysis as an effective and objective method for depression assessment using HTP drawings. The model not only improves accuracy but also offers potential applications in automated mental health screening. Future research should integrate multimodal data, such as speech and physiological signals, to enhance diagnostic precision.