Automated segmentation enables scalable image analysis, but capability differs by modality. Computed tomography (CT) has widely available multiclass models, whereas magnetic resonance imaging (MRI) beyond the brain remains constrained by limited tools and time-intensive pixel-wise labelling. Using machine-generated presegmentations that annotators refine can reduce effort. An approach assessed here applies CT-trained models to abdominal MRI using a simple inversion-based preprocessing step that adjusts appearance to resemble CT. Performance was examined on T1-weighted (T1w) and T2-weighted fat-saturated (T2wfs) sequences for a public multiclass model and a specialised renal tumour model. The central question was whether this method can yield anatomically plausible presegmentations that meaningfully speed annotation without training new cross-modal models.
Inversion-Based Adaptation Bridges Modality Differences
Dense tissues appear bright on CT yet often dark on MRI, impeding direct transfer. To narrow this gap, MRI images were intensity-clipped, inverted and given a black background by setting lowest intensities to zero. Two CT-trained models were applied to both original and inverted MRI: a multiclass model (TotalSegmentator-fast, version 2.2.1) and an in-house nnU-Net for renal tumours predicting left kidney, right kidney and primary tumour.
The dataset comprised 100 T1w and 100 T2wfs abdominal MRI sequences from 100 patients with histopathologically confirmed clear cell renal cell carcinoma. Reference labels spanned 24 abdominal structures from an automatic MRI model, while tumours and kidneys were manually annotated then reviewed by two radiologists, with tumour-free kidneys segmented automatically. Segmentation performance used the Dice similarity coefficient (DSC). Normality was assessed with Shapiro–Wilk, pairwise comparisons with Wilcoxon signed-rank tests, and correlations with Spearman’s rank, with Benjamini–Hochberg correction.
Must Read: DL Boosts MRI Efficiency in Postoperative Imaging
Inversion markedly improved how CT models interpreted T1w MRI. For the multiclass model, inverting and enforcing a black background converted largely failed outputs into anatomically plausible presegmentations across many classes. The renal tumour model also benefited on T1w after inversion, producing usable tumour and kidney masks suitable for refinement. By contrast, inversion did not consistently help T2wfs, indicating sequence-dependent effects.
Sequence-Specific Performance Highlights Where Gains Are Realised
Without preprocessing, the multiclass model was ineffective on T1w, detecting only the colon with a DSC of 0.38. On unprocessed T2wfs it segmented eight large organs with DSC above 0.40, including right kidney 0.60, spleen 0.55 and small bowel 0.55, but struggled with vessels and muscles, with the aorta at 0.17 and right iliopsoas at 0.19. After inversion with a black background, T1w performance improved broadly: twenty classes reached DSC between 0.40 and 0.77, with the right kidney at 0.77. The weakest T1w results after inversion were for structures dominated by air or fluid, including left and right lungs at 0.15 and 0.16, the gallbladder at 0.23 and vertebrae at 0.26. For T2wfs, inversion decreased performance for most classes, although the aorta improved from 0.17 to 0.30. Plain inversion without blackening the background underperformed compared with inversion plus background blackening.
The renal tumour model showed a similar pattern. On unprocessed T1w it failed to segment tumour, adjacent-normal kidney or contralateral-normal kidney, with DSC values at or below 0.03. On unprocessed T2wfs it roughly segmented kidneys with DSC 0.57 for adjacent-normal and 0.63 for contralateral-normal, but tumour segmentation was poor at 0.12. Inverting T1w and setting the background to black increased DSC to 0.71 for adjacent-normal kidney, 0.76 for contralateral-normal and 0.42 for tumour, creating presegmentations that annotators can refine. The same inversion prevented reliable segmentation on T2wfs, where all classes fell to 0.10 or below. Pairwise comparisons across preprocessing steps were significant at p < 0.001, apart from the comparison between inverted and inverted with black background for tumour prediction in T2wfs, where both failed and the difference was not significant.
Tumour volume affected T1w segmentation after inversion. Tumours were correctly localised in 75 scans, incorrectly in 19 and not detected in 6. Median tumour volumes in these groups were 35 cm³, 23 cm³ and 6 cm³ respectively. Tumours smaller than the overall median of 29 cm³ achieved a DSC of 0.22, whereas larger tumours reached 0.62, with a positive correlation between tumour volume and DSC.
Practical Considerations, Scope and Limitations
The inversion step appears especially valuable for T1w sequences and less suited to T2wfs, where many organs are already bright relative to surrounding tissues. Maintaining a black background improved segmentation, likely by enhancing contrast between anatomy and air. For musculoskeletal anatomy in T1w, improvements were observed for autochthonous and iliopsoas muscles and the spine. No evidence was presented for brain applications, and MRI-specific brain methods are established, indicating limited value for CT-to-MRI transfer in that region.
This approach targets presegmentation rather than definitive accuracy. It can provide a rapid first pass within active learning workflows, enabling annotators to prioritise high-quality cases and defer difficult ones. For certain sequences, CT models may be usable without augmentation, while for T1w, inversion with a black background can be decisive. Compared with machine learning-based image translation, inversion is simple, avoids synthetic image generation and the risk of hallucinated features, and may suit projects without resources to train cross-modal models or maintain interactive annotation software.
Several constraints frame interpretation. The reference standard combined automatic segmentation for multiple abdominal structures with manual tumour and kidney annotations, which affects the absolute meaning of DSC but is adequate for presegmentation aims. The focus on primary renal tumours excluded secondary lesions and metastases that may be clinically relevant yet fall outside the evaluation pipeline. The dataset was retrospective from a single institution, and MRI acquisition parameters varied, so generalisability beyond this setting is not established. External validation across centres would better define robustness across scanners and protocols.
Using CT-trained models with inversion-based preprocessing can generate MRI presegmentations that reduce manual effort, particularly for T1w abdominal imaging. Gains are sequence-dependent, with clear improvements for T1w after inversion and mixed effects in T2wfs. For renal tumours, inversion enabled localisation and segmentation of larger lesions on T1w, offering a workable starting point for annotation. Within workflows that value speed and scalability, this simple adaptation allows reuse of existing CT models for MRI while recognising scope limits, the need for careful review and the importance of broader validation.
Source: European Radiology Experimental
Image Credit: iStock
