Title: Comprehensive Evaluation of Intraoperative Ultrasound-to-MRI Synthesis for Brain Tumour Surgery

Intraoperative ultrasound (ioUS) serves as a practical and economical tool during brain tumour operations; however, its clinical utility is often hindered by complex interpretation challenges. These difficulties arise from non-standard acquisition planes, unique modality-specific artefacts, and the stark visual contrast between ioUS images and the preoperative MRI scans that underpin surgical planning, segmentation algorithms, and surgical expertise. Generating MRI-synthetic images from ioUS data offers a solution, allowing existing MRI-dependent infrastructure to be utilized during surgery without the need for additional scans.

Despite the potential of this approach, previous research has largely focused on evaluating individual architectures in isolation. To date, no comprehensive benchmark has compared diverse architectural paradigms, inference methods, and downstream clinical endpoints under a unified protocol. This study fills that void by analyzing the public ReMIND dataset, which comprises data from 76 patients, including 153 paired ioUS/T2w studies and 104 paired ioUS/FLAIR studies, with a patient-level split of 60 for training and 16 for held-out testing.

The study assessed six distinct image generation models: four GAN-based baselines (Pix2Pix, SwinPix2Pix, CycleGAN, and CUT), the transformer-enhanced ResViT, and the few-step diffusion model SynDiff. Each generator was subjected to four inference regimes (2D, 2.5D, 2D combined with 3D-refinement, and full-3D) and two target configurations (T2w-only and multi-task T2w + FLAIR), resulting in a total of 48 experimental configurations.

Performance was evaluated using standard image-fidelity metrics—SSIM, PSIM, MAE, and LPIPS—supplemented by a downstream segmentation task using nnU-Net v2 to identify tumours and resection cavities. The analysis also included subgroup breakdowns based on histological grade and whether the patient underwent reoperation.

The results indicated that no single architecture outperformed others across all metrics. Crucially, the study found that perceptual quality correlated most strongly with downstream clinical utility (LPIPS, r=-0.66, p<0.001). Conversely, higher SSIM scores were paradoxically linked to poorer utility (r=-0.64, p<0.001). Among the tested models, SynDiff using the 2.5D inference regime achieved the best preservation of downstream segmentation performance (U_Dice=0.55).

These findings suggest that global SSIM should be reported alongside, or preferably superseded by, perceptual and downstream-task metrics. Furthermore, the selection of an appropriate architecture should be tailored to specific factors, including the surgical phase, patient history, and the particular clinical objective at hand.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC