Deep learning study reveals key principles for TCM tongue diagnosis

Researchers have conducted a comprehensive ablation study to understand the effectiveness of deep learning models for traditional Chinese medicine tongue diagnosis. The study evaluated over 20 model versions using the TongueDx2 dataset and a merged dataset, comparing various backbone architectures, loss functions, augmentation strategies, and training methods. Key findings indicate that ConvNeXt-Tiny is the most parameter-efficient backbone, BCE loss outperforms Asymmetric Loss, and restrained color augmentation is crucial for performance. The research also highlighted the benefits of weak-group ensemble replacement over probability averaging and the significant impact of data scaling, while warning against expanding label dimensions which can lead to catastrophic model collapse. AI

IMPACT Identifies optimal deep learning configurations for medical image classification, potentially improving diagnostic tools.

RANK_REASON The cluster contains a research paper detailing an ablation study on deep learning models for a specific application. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• ConvNeXt-Tiny
• Hugging Face
• Tongue diagnosis
• TongueDx2
• traditional Chinese medicine

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →