What happens to a paper after it is accepted

Since 2021, MICCAI has published peer reviews for every accepted paper, but this information has not been analyzed before by "joining" it with the list of accepted papers by presentation type (oral, spotlight, or poster-only). Doing so enables us to answer interesting questions about what in the reviews, if anything, separates a talk from a poster: review scores, reviewer confidence, early acceptance, and subject area. 377 talks (orals and spotlights combined) were matched to 3,717 accepted papers across 5 years, which is 10.1% of everything accepted.

How many papers get a talk, and in which type?

Although spotlight talks started only in 2024, the share of accepted papers that received a podium presentation slot (oral or spotlight) has roughly remained consistent, even though the conference has grown in size from 531 to 1027 papers.

Mean review score by presentation type

Since the review score scale has changed over the years, the bar heights here show the mean score as a percentage of each year's own scale, and the whiskers show 95% bootstrap confidence intervals. Hover over each bar for more information, including standard deviation and group sizes.

Effect sizes

Cliff's δ captures how much more often papers from one group outscore papers from another. It is well suited for ordinal review scores, and stays comparable across years whose scales differ. In all years, oral papers outscored poster-only papers (δ from +0.28 to +0.74, every interval clear of zero). The gap is much smaller in 2025 (small) than in the years before it. The shaded bands in the plot mark the conventional thresholds for negligible (inner band; |δ| < 0.147) and small (outer band; |δ| < 0.33) effects. And in cases where the bars cross a shaded band, the difference is too small to matter.

Early-accepted papers are far more likely to get a talk

Early accepts are decided before the rebuttal round, based only on the first round of reviews. Perhaps consequently, early-accepted papers are heavily over-represented among orals in all years. The bars below show the share of early-accepted papers within each presentation type, and whiskers show 95% Wilson intervals. Hover over the bars for more information.

Is the score gap just early acceptance in disguise?

Comparing the scores of oral and poster papers within each acceptance group (early-accepted and not-early-accepted) shows that the gap between the two presentation types persists even after controlling for early acceptance: orals score higher than posters in both groups.

Confidence does not separate the types

Reviewers rate their own confidence on a four-level scale. Orals and posters receive very similar confidence ratings (average of about 3.3 every year), and which type gets (slightly) higher confidence ratings varies by year. So, reviewer confidence tells us nothing about why some papers are chosen for talks. An interesting deviation is the higher-than-expected confidence for spotlight papers in 2024, but this is only 18 papers, so it is hard to draw any concrete conclusions.

Selection rate by subject area, against the overall baseline

Use the Year dropdown to switch between a single year and all years pooled. The baseline is recomputed for whichever window you pick, because the share of accepted papers given a talk is not the same every year. Only areas with at least 25 accepted papers in the window are shown, and the most extreme are kept at each end. Note that a paper can carry several subject areas, so these groups overlap and do not sum to the overall rate. It is also worth remembering that MICCAI renames some of its subject areas between years, so in the pooled view, the hover text lists the years in which that subject area label was used.

Do talks come with code more often?

Not really, but recent trends are encouraging. Since 2024, the fraction of oral papers with code has both increased and is larger than the fraction for poster-only papers. The bars below show the year-by-year numbers and whiskers show 95% Wilson intervals.

Where do talks sit among all papers?

Here, we visualize the same specter2 model's embeddings used elsewhere on the site, in a umap projection. Poster-only papers are faded to background, and talks are drawn on top. As expected, orals and spotlights are scattered across the semantic map, but some areas appear to be favored for talks. Scroll to zoom, drag to pan, click a point to open the paper, and double click to reset the view.