Table 9.
Ablation experiment results.
| Model | PP-LCNet | MCA | Shape -IoU |
Map@0.5 (95% CI) | mAP@0.95 (95% CI) | P | R | Wsz | Latency |
|---|---|---|---|---|---|---|---|---|---|
| YOLOv10 baseline | – | – | – | 95.61% (95.30–95.93) | 74.77% (74.09–75.47) | 92.17% | 92.18% | 5.51 MB | 35.71 ms |
| YOLOv10 + PP-LCNet | ✓ | – | – | 97.17% (96.93–97.41) | 79.12% (78.55–79.69) | 97.28% | 93.38% | 5.37 MB | 32.47 ms |
| YOLOv10 + MCA | – | ✓ | – | 96.43% (96.17–96.68) | 78.02% (77.45–78.59) | 94.21% | 91.95% | 5.50 MB | 33.60 ms |
| YOLOv10 + shape-IoU | – | – | ✓ | 96.58% (96.32–96.83) | 78.59% (78.02–79.17) | 96.04% | 89.57% | 5.50 MB | 33.67 ms |
| Full model (all modules) | ✓ | ✓ | ✓ | 97.69% (97.46–97.91) | 79.39% (78.83–79.95) | 94.41% | 95.65% | 5.37 MB | 32.02 ms |
*95% CI calculation: we performed patient-level stratified 10-fold cross-validation. For each metric, the values obtained on the 10 validation folds were combined using a t-distribution (df = 9) to compute the mean ± t₉,0.975 · SD/√10.
Values shown in bold are from the Full model with all modules enabled.