MSCNet

A prostate multiparametric MRI examination is read as a set, not as a stack of pictures. T2-weighted imaging shows zonal anatomy and gland margins, diffusion-weighted imaging and its apparent-diffusion-coefficient map report cellularity, T1-weighted imaging covers haemorrhage and pelvic background, and PI-RADS scoring assumes all of them are present and interpretable. When a referring site omits a sequence, or susceptibility distortion makes the DWI locally non-diagnostic, the radiologist is left choosing between an uncertain read and a repeat scan.

MSCNet, developed at Nanyang Technological University with the radiology department of Ren Ji Hospital and posted as a preprint in August 2026, treats that as a routing problem rather than an image-translation one. Each available contrast keeps its own residual encoder, and at every feature scale a gating network conditioned on which target is being requested emits modality- and location-dependent weights that decide how much of each encoder's features enter the fused representation. A transformer bottleneck supplies global anatomical context, an attention-gated decoder suppresses features dominated by artefact, and an edge branch compares image gradients so that capsule and lesion boundaries survive reconstruction.

The same machinery does a second job: artefact patterns harvested from a clinical archive — susceptibility, motion ghosting, banding, Gibbs ringing and chemical shift — were transferred onto quality-controlled targets, so a sequence that is present but locally unreliable is repaired rather than replaced.

#Key Features

• Target-conditioned cross-modal gating: Sigmoid gates at five feature scales weight each available sequence by modality and spatial location, so an ADC map contributes differently when the requested target is DWI than when it is T2W.
• Boundary-aware supervision: An edge-gradient branch plus deep supervision at intermediate decoder scales preserve gland contours; removing the edge branch drops aggregate SSIM from 0.931 to 0.917, and removing the bottleneck to 0.899.
• Shared-routing variant: MSCNet-Shared covers all ten input-to-target configurations from one checkpoint using a binary availability mask, a learned target token and training-time modality dropout.
• Reconstruction-risk gating: A voxel-wise risk map, fitted on development data only and fixed before evaluation, flags unreliable outputs — failure-detection AUROC is 0.92 for sequence mismatch and 0.90 for lesion reconstruction error.
• Locked multicentre transfer: Weights, preprocessing rules and the risk threshold were frozen before an external cohort of 186 cases across three hospitals and three scanner vendors, with no site-specific retraining or recalibration.

#Technical Details

Task-specific MSCNet holds 154.1 million parameters and trains with AdamW under a composite objective summing Charbonnier, SSIM, MS-SSIM, perceptual, edge, wavelet, frequency and deep-supervision terms. Training drew on the public PI-CAI cohort (1,476 cases) and PROSTATEx (346 subjects) alongside institutional three- and four-sequence cohorts of 3,189 and 1,595 patients, plus 262,306 DICOM slices from 182 patients with clinically observed artefacts. Across the ten completion tasks the macro-average SSIM was 0.818 against 0.798 for the strongest task-matched comparator, with group means of 0.867 for the physically coupled DWI–ADC tasks, 0.727 for cross-contrast T2W synthesis and 0.796 for T1W completion. MSCNet-Shared reaches 0.804 using 158.3 million parameters in place of ten separate models. On the pooled external cohort, SSIM was 0.791, LPIPS 0.124 and diagnostic AUC 0.823; in a same-patient repeat-scan analysis of 204 paired acquisitions, restoration lifted SSIM from 0.714 to 0.883.

#Applications

The intended use is rescuing an examination rather than manufacturing training data: completing a referral study that arrived without DWI, salvaging an acquisition wrecked by motion, and avoiding an unnecessary repeat scan. In a blinded 1,000-case reader study with three radiologists, overall image quality met the prespecified −0.5-point non-inferiority margin against acquired images for DWI, ADC and T2W completion, but not for T1W. PI-RADS agreement with reference readings was 77.8% exact and 95.6% within one category, median reading time fell from 5.1 to 3.4 minutes per case, and 76% of cases were rated diagnostic-ready. In a separate 200-case cohort with biopsy or clinical follow-up, AUC for clinically significant cancer was 0.841 on MSCNet images against 0.860 on acquired images and 0.797 on the strongest baseline.

#Impact

Where catalog conditional generators such as CXR-DDPM and Lung-DDPM synthesize images to enlarge a training set, MSCNet generates into a named patient's examination, and the evaluation is built around that higher bar: radiologist adjudication of safety events, pathology-anchored discrimination, and a locked external cohort instead of a held-out fold. The authors keep the claim deliberately narrow. Neither code nor weights are public; a versioned repository is promised for the published record. Generated ADC maps should not be used for quantitative ADC measurement without separate calibration, transporting the locked risk threshold to the external hospitals retained only 48.9% of cases, and zero-shot performance on unseen sequence combinations was placed outside the principal claim. As a retrospectively validated preprint awaiting peer review, MSCNet is positioned as a reliability-gated adjunct that stays linked to the acquired examination.