Uni-Loc draws 3D CT lesion maps from scan-level labels, skipping voxel-by-voxel annotation
Synopsis
Localizing pathology in 3D CT normally requires experts to outline lesions voxel by voxel, which is costly and hard to extend across findings. Uni-Loc replaces the classification head with similarity scores between a pathology-conditioned query and spatial image features, so the same scores yield both a scan-level diagnosis and a 3D localization map, with gradient-guided thresholding outlining lesion regions. Across five localization tasks covering 355 annotated CT scans, it reaches macro-averaged voxel-level AUROC/AUPR of 96.1%/34.5%, against 80.7%/13.9% for relevance propagation and 88.7%/8.1% for gradient saliency.
Interpretation
A 3D CT model can now produce both a diagnostic call and a map of where the pathology sits in the volume while training only on scan-level diagnostic labels. BERT-style vision-language classifiers previously produced scan-level predictions from global representations through a separate classification head, without explicitly enforcing alignment between the predicted pathology and spatial features; pooling methods used in weakly supervised localization also assumed a fixed label set. On five localization tasks spanning four public chest CT datasets and 355 scans with voxel-level annotations, Uni-Loc reaches macro-averaged voxel-level AUROC/AUPR of 96.1%/34.5%, versus 80.7%/13.9% for Uni-CLS with relevance propagation and 88.7%/8.1% for gradient saliency.
Classification and localization share one set of spatial similarity scores: a pathology query is compared with image features at each spatial location, and those scores both form the 3D localization map and are pooled by attention weights into the scan-level diagnostic score. Concretely, the linear classification head on the [CLS] embedding in Uniferum is replaced by a similarity formulation, and a UNet-style progressive upsampler restores vision tokens, whose spatial resolution is reduced 32-fold, to voxel level. After gradient-guided thresholding, the localization maps reach a macro-averaged Dice of 0.310, above 0.256 for a fixed threshold; on CT-RATE and RAD-ChestCT, classification AUROC improves by 0.85 and 0.64 percentage points over Uni-CLS trained with the same recipe.
Perspective
Evaluation covers five pathologies across four public chest CT datasets, 355 cases with voxel-level annotations, with localization maps resampled to isotropic resolution and aligned to the reference volume; classification is evaluated on CT-RATE and the external RAD-ChestCT. It suits researchers and engineers working on weakly supervised medical image localization and vision-language models.
Localization remains difficult for some abnormalities, for example CLIP-style cosine similarity has higher AUROC than Uni-Loc on pleural effusion; gradient-guided thresholding gives slightly lower Dice than a fixed threshold on pleural effusion, mediastinal lymphadenopathy and atelectasis, with its macro-average advantage coming mainly from ground-glass opacity and infectious opacity. Removing the offset term in the similarity formulation significantly degrades voxel-level AUROC across pathologies, while AUPR changes are not significant. In classification comparisons, the CLIP-style baseline is only limitedly directly comparable because of differing evaluation settings.