# Report notes (seed 16)

Write your narrative here. `make_report.py` drops each `## section` into the
matching slot of report.pdf; any section you leave empty shows up as a yellow
TODO box in the PDF, so nothing can be silently forgotten.

Keep every claim consistent with the generated tables — you must be able to
defend each one in the viva.

## roll number

2026XXXX16

## a1 discussion

<!-- Acquisition marks, borders, cropping differences, class imbalance.
     Point at specific image IDs from A5_seed_selected_examples_seed16.png.
     Note the 299x299 images vs 256x256 masks and how you handled it. -->

## a2 ablation

<!-- Which single preprocessing change you ran and why it is clinically
     meaningful; the effect on each model (cite the comparison table). -->

## a2 errors

<!-- Eight errors per model. For each: image ID, true label, predicted label,
     confidence, and the explanation. The `evidence` column of
     Results/metrics/errors_*.csv gives you the supporting numbers. -->

## a2 conclusion

<!-- Which model won, and did chest-X-ray pretraining help? -->

## a3 discussion

<!-- Where segmentation fails; whether Dice differs across the four classes
     (cite the ANOVA); what the boundary metrics add over Dice. -->

## a4 discussion

<!-- Does attention behaviour match the classification metrics? Shortcut
     learning? Difference between correct and incorrect predictions? -->

## b2 discussion

<!-- How the EDA drove input size, augmentation and evaluation choices. -->

## b3 ablation

<!-- The imbalance targeted; effect of focal loss on recall, precision, FN, FP,
     mAP and small-polyp recall. -->

## b4 failures

<!-- Dominant failure patterns, each tied to an exploratory finding. -->

## analysis tasks

<!-- Which task was most preprocessing-sensitive; which metric most informative. -->

## analysis models

<!-- Metrics + Grad-CAM overlap + ablation + errors -> one verdict. -->

## limitations

<!-- Dataset quality, annotation quality, split design, generalisation. -->