Hybrid GAN-LSTM Framework for Diabetic Foot Ulcer Image Synthesis and Automated Diagnosis

Abinaya Vina¹Prajasree G¹Siddharth V¹Suresh Sankaranarayanan^2*Meenakshi K^1*Abdul Raouf Khan²Sharmila Sheik Banu Imam²Abdul Rahaman Wahab Sait³

• ¹SRM Institute of Science and Technology (Deemed to be University), Kattankulathur, India
• ²Department of Computer Science, King Faisal University, Al Hofuf, Saudi Arabia
• ³Department of Documents and Archives, King Faisal University,Al Ahsa, KSA, Al Hofuf, Saudi Arabia

The application of artificial intelligence in medical image analysis, particularly for Diabetic Foot Ulcers (DFUs), faces challenges due to insufficient well-labelled datasets for training diagnostic models. To address this, three hybrid models were developed, each integrating a generative component with a classification system. These models differ in classification structures, allowing for the assessment of generative data augmentation across various diagnostic formats. The models use advanced networks to produce high-quality synthetic DFU images, focusing on high-level generative stability and clinically relevant image accuracy. The three models vary in their generator architectures: the baseline CNN–LSTM for efficient spatial modeling, the EfficientNetV2M–LSTM for high-capacity feature extraction, and the EfficientNetV2S–LSTM for a balance between efficiency and synthesis quality. They also differ in computational complexity and training stability. The WGAN-GP + LSTM model ensures stable generative training through a critic network that evaluates synthetic images, promoting image diversity and reducing repetitive features. Utilizing a dataset of 5,894 clinically annotated DFU images from Lancashire Teaching Hospital, annotated by experienced professionals, the study demonstrates that synthetic images enhance disease classification accuracy and automated diagnostic systems' effectiveness. By maintaining clinically relevant variability in ulcer appearances, these images are crucial for real-world applications like screening, triage, and remote assessments, especially in underserved healthcare settings, paving the way for real-time clinical deployment.