Abstract

Quantitative FDG-PET/CT analysis depends on accurate tumour lesion delineation, but manual annotation is costly and variable, while fully supervised segmentation requires dense masks expensive to obtain at scale. We investigate weakly supervised FDG-PET lesion segmentation from patient-level labels alone, with no segmentation annotations used for classifier training or post-processing parameter tuning. Class activation mapping (CAM) provides lesion localisation heatmaps, but converting these continuous maps into binary masks remains a major challenge in annotation-free weakly supervised segmentation. To address this, we propose FLOWS (Fused Localisation-map Watershed Segmentation), an annotation-free and interpretable method for converting localisation heatmaps into binary lesion masks that replaces direct thresholding with a structured sequence of multi-scale hotspot detection, paired foreground/background seed generation, and marker-controlled watershed segmentation. FLOWS operates directly on CAM heatmaps while optionally incorporating PET intensity as an additional delineation cue. On the autoPET dataset, Layer-CAM variants provide the strongest localisation among evaluated CAM methods. Compared with Otsu thresholding, FLOWS improves pixel-wise Dice on correctly classified VGG16 scans from $42.0\%$ to $48.0\%$ using CAM alone, and performs similarly when fusing PET intensity, yielding a trade-off between recall and precision. Experimental results demonstrate that FLOWS can consistently reduce over-segmentation and false positives relative to direct heatmap thresholding, improving weakly supervised lesion segmentation without relying on ground-truth masks at any stage. Code is available at: https://github.com/an-scho/FLOWS.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMERGE_036.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=jOCb8do0bm

BibTex

@InProceedings{SchAnn_AnnotationFree_MICCAISAT2026,
        author = { Schoonjans, Anna AND De Maeyer, Roxanne AND De Ridder, Vicky AND DeKeyser, Simon AND Lazendić, Srđan},
        title = { { Annotation-Free Structured Delineation for Weakly Supervised PET Lesion Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17266},
        month = {pending},
        page = {pending}
}


back to top