The author developed a two-stream model for diabetic retinopathy staging using RetinaMNIST data, which consists of 1600 fundus images. The dataset is split into 1080 training, 120 validation, and 400 test images, categorized into five stages according to the ICDR scale.
The model architecture combines an EfficientNet-B2 vision encoder with a tabular branch processing four clinical indicators: HbA1c, diabetes duration, systolic blood pressure, and age. The author generated these clinical features synthetically based on the true disease stage, which introduced a data leakage.
Initial results showed a Quadratic Weighted Kappa (QWK) of 0.834. However, a control experiment using only images yielded a QWK of 0.713. A logistic regression model trained solely on the synthetic clinical features achieved a QWK of 0.816, confirming that the model was primarily relying on the leaked labels rather than the images.
The training process utilized AdamW optimization, AMP FP16, and Albumentations for image augmentation. Training the vision-only model on an NVIDIA Tesla T4 took 199.1 seconds, achieving a test QWK of 0.7129 and a Macro AUC of 0.8070. The author notes that the validation set is small and noisy, leading to unstable metrics.
The author concludes that clinical data should be validated by checking if a model trained only on tabular data performs similarly to the multimodal model. Furthermore, the author advises against using zero-input ablation, recommending instead the training of a separate baseline model to ensure accurate performance evaluation.
