arXiv Science⌕ Search

arXiv subjects

Bachir Taouli

Publications and source records attributed to Bachir Taouli.

4 recordsLinked to original sources

ORGANIC3D: Grounding 3D CT Vision Language Models in Organ Level Image Text Coupling

Vision language models (VLMs) have achieved strong performance in 3D computed tomography (CT) classification, but high accuracy does not reflect whether their predictions rely on the correct anatomical evidence. We introduce ORGANIC3D, a framework for evaluating and improving organ level image text alignment through controlled anatomical interventions. Our benchmark assesses 12 vision encoder checkpoints and 7 generative VLMs using matched patient comparisons, organ erasure, lesion replacement, and laterality perturbations, revealing substantial alignment failures even among top performing models. To learn organ level image text correspondence, ORGANIC3D aligns each finding sentence with the anatomy aware image features and trains on edited scans to anchor this alignment in the organ's image evidence. We further extend this framework to generative VLMs, promoting evidence dependent question answering and report generation. Trained exclusively on public data, ORGANIC3D achieves SOTA classification AUROC on CTRATE, Merlin, and the heldout RADChestCT dataset and reaches 77% laterality accuracy, compared with at most 66% for prior models. Its generative VLM matches or exceeds all evaluated publicly released VLMs on visual question answering, achieves the highest report accuracy, and retracts 96 to 99% of target organ findings after organ removal, versus a maximum of 69% for other models. Together, these results establish ORGANIC3D as a framework for testing whether 3D CT vision language models ground their predictions in anatomical evidence and for improving organ level alignment alongside predictive accuracy.

cs.CV↗

Detecting Early Kidney Allograft Fibrosis with Multi-b-value Spectral Diffusion MRI

Kidney allograft fibrosis is a marker of chronic kidney disease (CKD) and predicts functional decline, and eventual allograft failure. This study evaluates if spectral diffusion MRI can help detect early development and mild/moderate fibrosis in kidney allografts. In a prospective two-center study of kidney allografts, interstitial fibrosis and tubular atrophy (IFTA) was scored and eGFR was calculated from serum creatinine. Multi-b-value DWI (bvalues=[0,10,30,50,80,120,200,400,800mm2/s]) was post-processed with spectral diffusion, intravoxel incoherent motion (IVIM), and apparent diffusion coefficient (ADC). Connection between imaging parameters and biological processes was measured by Mann-Whitney U-test and Spearman's rank; diagnostic ability was measured by five-fold cross-validation univariate and multi-variate logistic regression. Quality control analyses included volunteer MRI (n=4) and inter-observer analysis (n=19). 99 patients were included (50$\pm$13yo, 64M/35F, 39 IFTA=0, 22 IFTA=2, 20 IFTA=4, 18 IFTA=6, 46 eGFR<=45mL/min/1.73m2, mean eGFR=47.5$\pm$21.3mL/min/1.73m2). Spectral diffusion detected fibrosis (IFTA>0) in patients with normal/stable eGFR>45ml/min/1.73m2 [AUC(95$\%$CI)=0.72(0.56,0.87),p=0.007]. Spectral diffusion detected mild/moderate fibrosis (IFTA=2-4) [AUC(95$\%$CI)=0.65(0.52,0.71),p=0.023], as did ADC [AUC(95$\%$CI)=0.71(0.54,0.87),p=0.013)]. eGFR, time-from-transplant, and allograft size could not. Interobserver correlation was >0.50 in 24 out of 40 diffusion parameters. Spectral diffusion MRI showed detection of mild/moderate fibrosis and fibrosis before decline in function. It is a promising method to detect early development of fibrosis and CKD before progression.

physics.med-ph↗

Estimation of Multi-Component Flow in the Kidney with Multi-b-value Spectral Diffusion

Purpose: Examine the theory and potential clinical application of estimated intravoxel flow of separated perfusion, tubular flow, and diffusion from multi-b-value DWI in kidney allografts. Methods: Multi-b-value DWI (9 b-values; 0-800 s/mm2) from a kidney cortex is simulated with anisotropic and non-Gaussian (i.e. anomalous) vascular, tubular, and tissue components and analyzed with a Bayesian biexponential, least-squares triexponential, and spectral diffusion MRI. Comparison and application of biexponential, triexponential, and spectral diffusion fD is demonstrated in a two-center study of 54 kidney allografts patients (21F/33M, 48.8 SD 10.5years) and compared to fibrosis (Banff 2017 interstitial fibrosis and tubular atrophy score 0-6 from clinical biopsies of the renal cortex), impaired kidney function (CKD-EPI 2021 eGFR<45ml/min/1.73m2), and proteinuria. Results: Spectral diffusion fD demonstrated strong correlation to input fD of the simulated anisotropic and anomalous components. It agreed with both three-component diffusion and two-component diffusion. fD showed similar or improved agreement and correlation to input compared to individual parameters, and similar or improved agreement to corresponding bi- and triexponential models. In kidney allografts, spectral diffusion fD showed higher allograft fibrosis score had higher fD_tissue, impaired allograft function showed reduced fD_tubule, and fD_vascular negatively correlated with proteinuria across diagnostic groups of function and fibrosis. Conclusions: Spectral diffusion MRI with multi-Gaussian fD as a flow proxy separated different anomalous and anisotropic diffusion components of perfusion, tubular flow, and tissue diffusion and may hold clinical value in diffusion MRI of kidney pathophysiology.

physics.med-ph↗

MRAnnotator: multi-Anatomy and many-Sequence MRI segmentation of 44 structures

In this retrospective study, we annotated 44 structures on two datasets: an internal dataset of 1,518 MRI sequences from 843 patients at the Mount Sinai Health System, and an external dataset of 397 MRI sequences from 263 patients for benchmarking. The internal dataset trained the nnU-Net model MRAnnotator, which demonstrated strong generalizability on the external dataset. MRAnnotator outperformed existing models such as TotalSegmentator MRI and MRSegmentator on both datasets, achieving an overall average Dice score of 0.878 on the internal dataset and 0.875 on the external set. Model weights are available on GitHub, and the external test set can be shared upon request.

eess.IV↗