Automatic Diagnosis of Cushing's Syndrome Using Combined Handcrafted and Deep Features from Patient Facial Images

Changwei Song, Jiaqi Qiang, Hongjun Liu, Jianqiang Li, Hui Pan, Qing Zhao, Jiuzuo Huang, Shi Chen

Keywords: Cushing's Syndrome, Foundation Models, Pre-trained Models, Transformer

Abstract: Cushing's syndrome is an endocrine and metabolic disorder caused by adrenal cortical hyperfunction leading to excessive glucocorticoid secretion, and its characteristic facial phenotypes (e.g., moon face, skin striae distensae) provide important objective evidence for clinical diagnosis. In recent years, deep learning-based computer-aided diagnostic techniques have achieved automatic recognition of facial features via pre-trained convolutional neural networks, but most existing methods fail to fully utilize the complementary information contained in handcrafted features guided by medical prior knowledge. To address this limitation, this study innovatively proposes an intelligent diagnostic model based on multi-view feature fusion, constructing a tripartite framework comprising deep feature extraction, handcrafted feature extraction, and feature fusion. First, a transfer learning strategy is adopted to realize hierarchical extraction of high-level semantic features using the pre-trained Transformer model (DINOv2). Second, combined with medical image feature engineering, a dual-channel handcrafted feature extractor is designed to parallelly extract HOG texture features and LBP local pattern features, effectively incorporating domain prior knowledge. Finally, a dynamic feature fusion module based on attention mechanism is developed to achieve adaptive fusion of deep features and handcrafted features through a learnable weight assignment strategy. Experimental results demonstrate that the proposed model significantly outperforms single-feature models in terms of accuracy (98.5\%) and F1-score (94.1\%). This study provides a new paradigm integrating medical prior knowledge and data-driven approaches for the intelligent auxiliary diagnosis of Cushing's syndrome, exhibiting promising clinical translation potential.

Submission Number: 52