| # Stage 1 Binary Detection — accuracy + recall oriented | |
| # Target: 90–95% overall accuracy with strong cancer sensitivity. | |
| model_name: "efficientnet_b0" | |
| pretrained: true | |
| in_channels: 3 | |
| num_classes: 1 | |
| img_size: 224 | |
| seed: 42 | |
| batch_size: 16 | |
| learning_rate: 0.0003 | |
| weight_decay: 0.0001 | |
| epochs: 14 | |
| # Train classifier head only for N epochs, then fine-tune full net | |
| freeze_epochs: 2 | |
| amp_enabled: true | |
| # Print progress every N train batches (CPU runs are long) | |
| log_every: 25 | |
| threshold: 0.35 | |
| # FN penalty to push recall higher on ambiguous cancer patches | |
| pos_weight_boost: 1.5 | |
| # Clip exploding gradients during full fine-tune | |
| grad_clip: 1.0 | |
| # Stop if val accuracy stalls (0 = disabled) | |
| early_stopping_patience: 5 | |
| device: "cpu" | |
| data: | |
| dataset_ids: | |
| - lc25000 | |
| - patch_camelyon | |
| - breast_histopathology_patches | |
| - skin_lesion_hm10000_binary_2k | |
| - camelyon17_jxie | |
| - camelyon17_djghosh | |
| # Increased sample count for multi-organ histology coverage | |
| max_per_dataset: 4000 | |
| balanced_sampling: true | |
| apply_stain_norm: false | |
| num_workers: 0 | |
| output: | |
| artifacts_dir: "models/stage1/artifacts" | |