This challenge aimed to classify peripheral blood smears into eight categories using deep learning. We explored:
- Custom CNNs
- Transfer Learning (ImageNet)
- Test Time Augmentation (TTA)
Best approach: Transfer Learning with ConvNeXt models.
Dataset: 13,758 images, with 1,799 duplicates removed. Maintaining the original class distribution provided the best results.
Key augmentations:
- Rotation, flipping
- Contrast, brightness adjustments
- Gaussian noise
- RandAugment pipeline
Performance Boost: VGG models improved from 0.61 → 0.72, ConvNeXt from 0.88 → 0.93.
Top models: ConvNeXt > EfficientNet > VGG.
- Transfer Learning: Used pre-trained ImageNet weights.
- TTA: Applied but limited by Codabench constraints.
- Hyperparameter Tuning: Optimized learning rate, batch size.
- Best optimizer: NAdam.
Final Results: Accuracy 0.9427, F1 0.9385.