Lung segmentation models such as R231 are widely used, but they fail on severe tuberculosis (TB) CT: cavities and consolidations look like non-lung tissue and are left out of the mask. We present LungMend, a post-processing model that corrects such masks using only healthy anatomy. A latent diffusion model, trained on healthy lungs with synthetically corrupted masks, learns to recover the complete lung from the faulty mask and the bone anatomy, without any pathological data or CT intensities. LungMend recovers cavities and consolidations that R231 misses and improves accuracy on expert-annotated slices. Code: https://github.com/noureddinekhiati/LungMend
Figures & tables
Figure 1: Overview of LungMend. Training (Stage 2): the SDT of a healthy lung mask is encoded by the frozen VAE encoder E∗ (pre-trained in Stage 1) into z0 . A pathology-mimicking corruption of the mask gives zcorr , and bone occupancy and distance give cbone . The U-Net V θ takes the noisy latent zt and [zcorr,cbone] (18 channels, 163 ) and predicts vθ . Inference: N samples are decoded by D∗ and averaged into the corrected mask and an uncertainty map. CFG denotes: classifier-free guidance.
Expert slices
Lesion recall
Method
Dice ↑
ASD ↓
HD95 ↓
Cav. ↑
Cons. ↑
R231 [ 1 ]
0.909
4.38
27.4
0.869
0.732
R231 + closing
0.929
3.79
22.9
0.934
0.868
Post-DAE [ 7 ]
0.768
11.87
36.5
0.968
0.942
Det. 3D U-Net
0.896
5.04
26.1
0.557
0.731
TotalSegmentator
0.919
4.29
24.7
0.951
0.723
Table 1: Results on the 200 expert slices of severe TB cases (mean; ASD and HD95 in mm). Best in bold.
Dice ↑
ASD ↓
A. Training components
Full recipe (reference)
0.964
2.34
Latent → voxel-space diffusion
0.917
4.54
SDT → binary mask
0.795
15.02
No Dice term in VAE
0.935
3.68
No bone mask
0.929
3.47
Table 2: Ablations on the 200 expert slices (ASD in mm). A : one training component removed at a time, with the model retrained using the same protocol as the full recipe; all rows use plain sampling (1 sample, w=1 ). B : inference steps added one at a time to the final model. R231 input: Dice 0.909.
Figure 2: Qualitative results on severe TB. (a) ImageCLEF TB axial slices: R231+closing (red, top) and LungMend (blue, bottom) masks overlaid on the CT, with the annotation in yellow. The region added by LungMend is green; the volume added over R231 is given under each case. (b) Avicenne TB: coronal slices and 3D renderings for R231 and LungMend. LungMend recovers cavities and consolidated lung that R231 excludes, and removes the stair-step fragmentation of the R231 3D surface.
Accurate quantification of lung disease severity from chest imaging is critical for clinical decision-making and resource allocation. We propose a tri-modal deep learning framework, TMF-RSE (Tri-Modal Fusion with Regional Semantics and Evidential Uncertainty), that combines appearance features from two-dimensional chest inputs, structural features from lung segmentation masks, and semantic features from vision-language models (VLMs) for severity quantification. Our approach employs complementary fusion mechanisms that integrate semantic guidance, structural priors, and hierarchical interactions across modalities. The model employs evidential regression to provide both severity predictions and uncertainty estimates. Experiments on the Per-COVID-19 CT and RALO datasets show that TMF-RSE outperforms recent transformer-based baselines, achieving MAE of 4.02 and Pearson correlation of 0.9629 on Per-COVID-19 validation, and 0.339 MAE / 0.973 PC on RALO geographic extent.
Fadi Abdeladhim Zidi, Salah Eddine Bekhouche, Abdellah Zakaria Sellam +3
Institute of Applied Sciences and Intelligent Systems (ISASI), CNR, 73100 Lecce, Italy · Department of Computer Science and Artificial Intelligence, University of the Basque Country (UPV/EHU), Spain · University of Salento, 73100 Lecce, Italy
Curating large labeled CT datasets is expensive and constrained by patient privacy. We present NV-Generate-BodyMask, a 3D latent diffusion model that synthesizes whole-body CT segmentation masks with precise, joint control over organ anatomy, tumor burden, and patient demographics. The model operates in the compressed latent space of a frozen mask autoencoder, generating 2563 volumetric label maps conditioned on a 19-dimensional vector via Adaptive Group Normalization and cross-attention. A size-adaptive ROI loss with spatial halo weighting and a paired tumor injection pipeline enable accurate tumor size conditioning despite the scarcity of annotated tumor cases. Trained on 8,571 masks across 26 cohorts spanning 132 labels, the model achieves Spearman ρ=0.70--0.92 for organ size, 85--100% tumor recall for lung, pancreatic, colon, and bone lesions, and strong demographic alignment (BMI ρ=0.78, sex-conditioned perfect prostate control).
Can Zhao, Pengfei Guo, Dong Yang +8
NVIDIA · National Institutes of Health · University of Arkansas for Medical Sciences
Delineating lung tumours on computed tomography (CT) takes a considerable share of the time spent on radiotherapy planning, and a contour proposed by a model can be refined interactively by the clinician. Promptable foundation models such as SAM 3 support this workflow by writing each correction into a session memory that conditions the remaining slices, while the model weights stay fixed. On 690 test cases from five public CT cohorts, fine-tuning SAM 3 on lung tumours raises the Dice obtained from a single point prompt from 0.298 to 0.757, and seven rounds of corrections raise it further to 0.765, but under memory conditioning alone the accuracy on slices the annotator has not touched stops improving after six rounds. We therefore treat each correction as a training signal and propose LeCor, which performs test-time training on a small set of case adapters that are reset for every case and meta-learned such that a single gradient step driven by a click improves the slices that were not clicked. On the 133 test cases that span at least eight slices, LeCor raises the Dice reached after seven correction rounds from 0.787 with the fine-tuned model to 0.827, reduces the number of cases that never reach a Dice of 0.80 from 47 to 27, and reaches in three correction rounds the accuracy that the fine-tuned model attains in seven.
Yi Luo, Yike Guo, Wenxuan Li +3
Department of Biomedical Engineering, Johns Hopkins University · Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University · Department of Computer Science, Johns Hopkins University +1