ICANSee

Enhanced Neoplasia Detection in Endoscopic Images with Domain-Specific Pretraining and Focal Loss for Severe Class Imbalance

ICANSee is a simple set of improvements over the original RARE Challenge baseline to reduce data leakage, prevent overfitting, and improve recall in a highly imbalanced setting.

What changed

• Evaluation protocol: validation-only by default, with optional separate test set via --eval_data_path.
• Early stopping: stops training when the chosen validation metric stops improving (patience/min_delta configurable).
• Sensitivity-focused training option: allow saving the best checkpoint by Sensitivity when recall is the priority.
• Focal loss tuning: support for focal_alpha/focal_gamma to bias towards the positive (neoplasia) class.
• Gastronet baseline integration: easy evaluation and use of pretrained weights (--use_gastronet or --pretrained_path).

How to run ICANSee

Train (validation-only evaluation, AUPRC as selection metric):

env PYTHONPATH=. python3 train.py \
  --data_path data \
  --model resnet50 \
  --epochs 30 \
  --loss focal_loss --focal_alpha 0.9 --focal_gamma 1.0 \
  --sampling oversample \
  --use_gastronet \
  --save_metric AUPRC \
  --patience 8 --min_delta 0.001

---

Focal loss

The focal loss $\mathrm{FL}$ with focusing parameter $\gamma \ge 0$ and class weight $\alpha \in [0,1]$ is $$\mathrm{FL}(p_t) = -, \alpha , (1 - p_t)^{\gamma} , \log(p_t).$$

Equivalently, written per-class in terms of $p$: $$\mathrm{FL}(y, p) = -, \alpha, y, (1-p)^{\gamma} \log(p) - (1-\alpha), (1-y), p^{\gamma} \log(1-p).$$

• $\gamma$: increases down-weighting of easy examples as it grows.
• $\alpha$: balances positive/negative classes; higher $\alpha$ emphasizes the positive class.