A computer engineering student is seeking advice on improving a 5-class diabetic retinopathy detection model trained on the APTOS 2019 dataset. The model exhibits inconsistent predictions, misclassifying classes like Moderate as Severe or Proliferative, and shows high confidence even with incorrect predictions, especially on images outside the training dataset. The student has already tried various pretrained models, preprocessing techniques, and test-time augmentation, and is now considering ensemble models or investigating domain shift, class imbalance, or preprocessing issues. AI
IMPACT This query highlights common challenges in medical image classification, such as domain shift and class imbalance, which are critical for deploying reliable AI systems in healthcare.
RANK_REASON User is asking for advice on an existing model, not announcing a new release or research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
