The mIoU Headline Flatters SSL: A Label-Efficient Head-to-Head of Four Self-Supervised\\ Pretraining Families on Lung-Tumor CT

• An isolated four-way SSL comparison on lung-tumor CT
segmentation, with the decoder, split, and labelled pool
identical across methods and the only free variable being
the pretext family.
• A standalone YOLOv26-Semantic C2f + PAN-FPN de-
coder re-implementation plus a small ViT feature adapter,
so the decoder can mount on either a ResNet-50 or a
plain ViT (single-scale token) encoder.
• A quantitative demonstration of the divergence between
mIoU and foreground IoU on 0.14%-foreground data: SSL
methods reach 78%–97% of the supervised mIoU while
reaching only 16.5%–52.2% of the supervised tumor IoU
on the same runs.