Inter-Slice Representation Outweighs Bounding-Box Supervision Extent in Lightweight 2.5D Pulmonary Nodule Detection: A Whole-Volume Benchmark on LUNA16.

Background/

Objectives

Lightweight detectors are attractive for high-throughput low-dose CT lung-cancer screening, yet the training-time choices governing their accuracy are usually fixed without justification, as is the protocol used to evaluate them. We benchmarked the inter-slice input representation and the extent of bounding-box supervision for 2.5D pulmonary nodule detection under the official LUNA16 protocol.

Methods

Using one YOLO11n backbone we compared a 2D central-slice baseline, thin-slab maximum-intensity projection, and adjacent-slice 2.5D input under loose and core-focused (60% of diameter) supervision. Evaluation followed the official subset0-subset9 ten-fold protocol over all 888 scans and 1186 nodules, with whole-volume inference across 227,225 axial slices, annotations_excluded.csv applied, and paired scan-level bootstrap intervals. Supervision ratio, minimum box size, negative mining, three seeds and a YOLO26n backbone were ablated.

Results

The representation dominated. Adjacent-slice input reached a competition performance metric (CPM) of 0.7795 (95% CI 0.7517-0.8005) against 0.6498 for the 2D baseline (+0.1297; 10 of 10 folds; d_z = 2.96), while thin-slab projection (0.5965) was worse than a plain 2D slice. Core-focused supervision gave no benefit (-0.0209) and no ratio improved on full-diameter supervision. Re-scoring the same checkpoints over only the slices holding an annotated nodule centre reversed the supervision result (+0.0027) and shrank the representation effect fourfold (+0.0348).

Conclusions

Inter-slice representation, not supervision extent, is the dominant design factor, and the slice set searched at inference decides whether either factor is measurable.