Synthetic MRI-Assisted and Self-Supervised Adaptive Segmentation of Organs-at-Risk (OARs) in MRI-Based Radiation Therapy

PURPOSE/OBJECTIVE(S)

This study proposes a self-supervised solution for OAR segmentation, combining patch-based adaptation and unsupervised synthesis of T2-weighted MRI data to finetune the segmentation model. The aim is to improve adaptation to patient anatomy, overcome limited annotated MRI data, and enhance the generalizability of automatic segmentation models for gynecological cancers.

MATERIALS/METHODS

The study used a patch-based cycle consistent generative adversarial network (cycle-GAN) for unsupervised MRI synthesis from CT scans of 20 patients, and a residual U-Net model for OARs segmentation. The segmentation model was trained and validated on synthetic MRI (sMRI) of 103 and 25 patient scans respectively, then finetuned on 78 MRI scans from radiation therapy fractions of 15 additional patients through three-fold cross validation. Self-supervised adaptation was applied, incorporating affine and elastic deformations, intensity shifting, and scaling. The model was trained on 96 × 96 × 96 sub-volumes and validated on entire pelvic sections of the same images. A combination of Dice and weighted cross entropy (CE) losses, with weights assigned for bladder (1), small bowel (1), rectum (2), sigmoid (2), left femoral head (0), and right femoral head (0), was used for OAR segmentation. The performance was evaluated against the model trained only on a limited number of acquired MRI data, as well as sMRI pretrained models with encoder weight freezing and either equal weighting or soft-tissue adjusted weighting.

RESULTS

Our sMRI-assisted approach showed improved performance for challenging pelvic OARs compared to the method using only the acquired MRI data. The self-supervised fraction-adaptive segmentation results indicated better performance in target soft-tissues when using at least one treatment fraction for organ-specific adaptation.

CONCLUSION

Our framework leverages pre-existing CT planning data for gynecological cancers to enhance the segmentation performance of OARs during MR-guided adaptive treatments. This approach offers substantial benefits for the radiation therapy workflow, including reduced variability in per-fraction segmentation and clinical burden. Further studies that involve human expert evaluations will be conducted to assess the practicality of this approach in radiation therapy.

Measurement reproducibility of perfusion fraction and pseudodiffusion coefficient derived by intravoxel incoherent motion diffusion-weighted MR imaging in normal liver and metastases

OBJECTIVE

To determine the measurement reproducibility of perfusion fraction f, pseudodiffusion coefficient D and diffusion coefficient D in colorectal liver metastases and normal liver.

METHODS

Fourteen patients with known colorectal liver metastases were examined twice using respiratory-triggered echo-planar DW-MRI with eight b values (0 to 900 s/mm(2)) 1 h apart. Regions of interests were drawn around target metastasis and normal liver in each patient to derive ADC (all b values), ADC(high) (b values ≥ 100 s/mm(2)) and intravoxel incoherent motion (IVIM) parameters f, D and D by least squares data fitting. Short-term measurement reproducibility of median ADC, ADC(high), f, D and D values were derived from Bland-Altman analysis.

RESULTS

The measurement reproducibility for ADC, ADC(high) and D was worst in colorectal liver metastases (-21 % to +25 %) compared with liver parenchyma (-6 % to +8 %). Poor measurement reproducibility was observed for the perfusion-sensitive parameters of f (-75 % to +241 %) and D (-89 % to +2,120 %) in metastases, and to a lesser extent the f (-24 % to +25 %) and D (-31 % to +59 %) of liver.

CONCLUSIONS

Estimates of f and D derived from the widely used least squares IVIM fitting showed poor measurement reproducibility. Efforts should be made to improve the measurement reproducibility of perfusion-sensitive IVIM parameters.

Refinement of local and long-range structural order in theophylline-binding RNA using (13)C-(1)H residual dipolar couplings and restrained molecular dynamics

13C-(1)H residual dipolar couplings (RDC) have been measured for the bases and sugars in the theophylline-binding RNA aptamer, dissolved in filamentous phage medium, and used to investigate the long-range structural and dynamic behavior of the molecule in the solution state. The orientation dependent RDC provide additional restraints to further refine the overall structure of the RNA-theophylline complex, whose long-range order was poorly defined in the NOE-based structural ensemble. Structure refinement using RDC normally assumes that molecular alignment can be characterized by a single tensor and that the molecule is essentially rigid. To address the validity of this assumption for the complex of interest, we have analyzed distinct domains of the RNA molecule separately, so that local structure and alignment tensors experienced by each region are independently determined. Alignment tensors for the stem regions of the molecule were allowed to float freely during a restrained molecular dynamics structure refinement protocol and found to converge to similar magnitudes. During the second stage of the calculation, a single alignment tensor was thus applied for the whole molecule and an average molecular conformation satisfying all experimental data was determined. Semirigid-body molecular dynamics calculations were used to reorient the refined helical regions to a relative orientation consistent with this alignment tensor, allowing determination of the global conformation of the molecule. Simultaneously, the local structure of the theophylline-binding core of the molecule was refined under the influence of this common tensor. The final ensemble has an average pairwise root mean square deviation of 1.50 +/- 0.19 A taken over all heavy atoms, compared to 3.5 +/- 1.1 A for the ensemble determined without residual dipolar coupling. This study illustrates the importance of considering both the local and long-range nature of RDC when applying these restraints to structure refinements of nucleic acids.

Solution structure of 2-(pyrido[1,2-e]purin-4-yl)amino-ethanol intercalated in the DNA duplex d(CGATCG)2

The solution structure of the complex formed between d(CGATCG)(2) and 2-(pyrido[1,2-e]purin-4-yl)amino-ethanol, a new antitumor drug under design, has been resolved using NMR spectroscopy and restrained molecular dynamic simulations. The drug molecule intercalates between each of the CpG dinucleotide steps with its side chain lying in the minor groove. Analysis of NMR data establishes a weak stacking interaction between the intercalated ligand and the DNA bases; however, the drug/DNA affinity is enhanced by a hydrogen bond between the hydroxyl group of the end of the intercalant side chain and the amide group of guanine G6. Unrestrained molecular dynamic simulations performed in a water box confirm the stability of the intercalation model. The structure of the intercalated complex enables insight into the structure-activity relationship, allowing rationalization of the design of new antineoplasic agents.

A novel interactive tool for rigid-body modeling of multi-domain macromolecules using residual dipolar couplings

Residual dipolar couplings (RDC), measured by dissolving proteins in dilute liquid crystal media, or by studying naturally paramagnetic molecules, have rapidly become established as routine measurements in the investigation of the structure of macromolecules by NMR. One of the most obvious applications of the previously inaccessible long-range angular information afforded by RDC is the accurate definition of domain orientation in multi-module macromolecules or complexes. In this paper we describe a novel program developed to allow the determination of alignment tensor parameters for individual or multiple domains in macromolecules from residual dipolar couplings and to facilitate their manipulation to construct low-resolution models of macromolecular structure. For multi-domain systems the program determines the relative orientation of individual structured domains, and provides graphical user-driven rigid-body modeling of the different modules relative to the common tensorial frame. Translational freedom in the common frame, and equivalent rotations about the diagonalized (x,y,z) axes are used to position the different modules in the common frame to find a model in best agreement with experimentally measured couplings alone or in combination with additional experimental or covalent information.

De novo determination of protein structure by NMR using orientational and long-range order restraints

Orientational and novel long-range order restraints available from paramagnetic systems have been used to determine the backbone solution structure of the cytochrome c' protein to atomic resolution in the complete absence of restraints derived from the nuclear Overhauser effect. By exploiting the complementary geometric dependence of paramagnetic pseudocontact shifts and the recently proposed Curie-dipolar cross correlated relaxation effect, in combination with orientational constraints derived from residual dipolar coupling, autorelaxation rate ratios and secondary structure constraints, it is possible to define uniquely the fold and refine the tertiary structure of the protein (0.73 A backbone rmsd for 82/129 amino acid residues) starting from random atomic Cartesian coordinates. The structure calculation protocol, developed using specific models to describe the novel constraint interactions, is robust, requiring no precise a priori estimation of the various interaction strengths, and provides unambiguous convergence based only on the value of the target function. Tensor eigenvalues and their component orientations are allowed to float freely, and are thus simultaneously determined, and found to converge, during the structure calculation.

Combined structural and biochemical analysis of the H-T complex in the glycine decarboxylase cycle: evidence for a destabilization mechanism of the H-protein

The lipoate containing H-protein plays a pivotal role in the catalytic cycle of the glycine decarboxylase complex (GDC), undergoing reducing methylamination, methylene transfer, and oxidation. The transfer of the CH(2) group is catalyzed by the T-protein, which forms a 1:1 complex with the methylamine-loaded H-protein (Hmet). The methylamine group is then deaminated and transferred to the tetrahydrofolate-polyglutamate (H(4)FGlu(n)) cofactor of T-protein, forming methylenetetrahydrofolate-polyglutamate. The methylamine group is buried inside the protein structure and highly stable. Experimental data show that the H(4)FGlu(n) alone does not induce transfer of the methylene group, and molecular modeling also indicates that the reaction cannot take place without significant structural perturbations of the H-protein. We have, therefore, investigated the effect of the presence of the T-protein on the stability of Hmet. Addition of T-protein without H(4)FGlu(n) greatly increases the rate of the unloading reaction of Hmet, reducing the activation energy by about 20 kcal mol(-1). Differences of the (1)H and (15)N chemical shifts of the H-protein in its isolated form and in the complex with the T-protein show that the interaction surface for the H-protein is localized on one side of the cleft where the lipoate arm is positioned. This suggests that the role of the T-protein is not only to locate the tetrahydrofolate cofactor in a position favorable for a nucleophilic attack on the methylene carbon but also to destabilize the H-protein in order to facilitate the unlocking of the arm and initiate the reaction.

Simultaneous determination of disulphide bridge topology and three-dimensional structure using ambiguous intersulphur distance restraints: possibilities and limitations

Knowledge of the native disulphide bridge topology allows the introduction of conformational restraints between remote parts of the peptide chain. This information is therefore of great importance for the successful determination of the three-dimensional structure of cysteine-rich proteins by NMR spectroscopy. In this paper we investigate the limitations of using ambiguous intersulphur restraints [Nilges, M. (1995) J. Mol. Biol., 245, 645-660] associated with NMR experimental information to determine the native disulphide bridge pattern. Using these restraints in a simulated annealing protocol we have determined the correct topology of numerous examples, including a protein with seven disulphide bridges (phospholipase A2) and a protein in which 25% of the total number of residues are cysteines (mu-conotoxin GIIIB). We have also characterised the behaviour of the method when only limited experimental data is available, and find that the proposed protocol permits disulphide bridge determination even with a small number of restraints (around 5 NOEs--including a long-range restraint--per residue). In addition, we have shown that under these conditions the use of a reduced penalty function allows the identification of misassigned NOE restraints. These results indicate that the use of ambiguous intersulphur distances with the proposed simulated annealing protocol is a general method for the determination of disulphide bridge topology, particularly interesting in the first steps of NMR study of cysteine-rich proteins. Comparison with previously proposed protocols indicates that the presented method is more reliable and the interpretation of results is straightforward.

Efficient analysis of macromolecular rotational diffusion from heteronuclear relaxation data

A novel program has been developed for the interpretation of 15N relaxation rates in terms of macromolecular anisotropic rotational diffusion. The program is based on a highly efficient simulated annealing/minimization algorithm, designed specifically to search the parametric space described by the isotropic, axially symmetric and fully anisotropic rotational diffusion tensor models. The high efficiency of this algorithm allows extensive noise-based Monte Carlo error analysis. Relevant statistical tests are systematically applied to provide confidence limits for the proposed tensorial models. The program is illustrated here using the example of the cytochrome c' from Rhodobacter capsulatus, a four-helix bundle heme protein, for which data at three different field strengths were independently analysed and compared.

Investigation of the local structure and dynamics of the H subunit of the mitochondrial glycine decarboxylase using heteronuclear NMR spectroscopy

The lipoate-dependent H protein plays a pivotal role in the catalytic cycle of the glycine decarboxylase complex (GDC), undergoing reducing methylamination, methylene transfer, and oxidation. The local structure and backbone dynamics of the methylamine-loaded H (Hmet), oxidized H (Hox), and H apoprotein (Hapo) have been investigated in solution. Filtered NOESY experiments using a [13C]Hmet as well as comparison of the heteronuclear shifts between the Hox and Hmet proteins demonstrate that the methylamine group is located inside a cleft of the protein. Furthermore, this group appears to be locked in this configuration as indicated by the high value of the activation energy (37 kcal/mol) of the global unloading reaction and by its restricted mobility, deduced from 13C relaxation measurements. Comparisons of the 1H and 15N chemical shifts and 15N relaxation in the three forms suggest that part of the lipoyl-lysine arm interacts with the protein polypeptide in the Hox and Hmet. The major change induced by the loading of the methylamine group concerns the C-terminal helix whose mobility becomes completely restricted compared to those of the Hox and Hapo. This C-terminal helix exhibits different reorientational characteristics in the three forms, which can be explained in the Hapo by a model consisting of a twisting motion about an axis passing through the helix. Our results indicate that the model of a freely swinging arm proposed for other lipoate-containing proteins is not acceptable in solution for the GDC. The implication of this observation in terms of the mechanism of the interaction of the H protein with the T protein, its physiological partner during the catalytic cycle, is discussed.

A structural homologue of colipase in black mamba venom revealed by NMR floating disulphide bridge analysis

The solution structure of mamba intestinal toxin 1 (MIT1), isolated from Dendroaspis polylepis polylepis venom, has been determined. This molecule is a cysteine-rich polypeptide exhibiting no recognised family membership. Resistance to MIT1 to classical specific endoproteases produced contradictory NMR and biochemical information concerning disulphide-bridge topology. We have used distance restraints allowing ambiguous partners between S atoms in combination with NMR-derived structural information, to correctly determine the disulphide-bridge topology. The resultant solution structure of MIT1, determined to a resolution of 0.5 A, reveals an unexpectedly similar global fold with respect to colipase, a protein involved in fatty acid digestion. Colipase exhibits an analogous resistance to endoprotease activity, indicating for the first time the possible topological origins of this biochemical property. The biochemical and structural homology permitted us to propose a mechanically related digestive function for MIT1 and provides novel information concerning snake venom protein evolution.

Solution structure, rotational diffusion anisotropy and local backbone dynamics of Rhodobacter capsulatus cytochrome c2

The solution structure, backbone dynamics and rotational diffusion of the Rhodobacter capsulatus cytochrome c2 have been determined using heteronuclear NMR spectroscopy. In all, 1204 NOE-derived distances were used in the structure calculation to give a final ensemble with 0.59(+/-0.08) A rms deviation for the backbone atoms (C, Calpha and N) with respect to the mean coordinates. There is no major difference between the solution structure and the previously solved X-ray crystal structure (1.07(+/-0.07) A rms difference for the backbone atoms), although certain significant local structural differences have been identified. This protein contains five helical regions and a histidine-heme binding domain, connected by a series of structured loops. The orientation of the helices provides an excellent sampling of angular space and thus allows a precise characterization of the anisotropic diffusion tensor. Analysis of the hydrodynamics of the protein has been performed by interpretation of the 15N relaxation data using isotropic, axially asymmetric and fully anisotropic diffusion tensors. The protein can be shown to exhibit significant anisotropic reorientation with a diffusion tensor with principal axes values of 1.405(+/-0.031)x10(7) s-1, 1.566(+/-0.051)x10(7) s-1 and 1.829(+/-0.054)x10(7) s-1. Hydrodynamic calculations performed on the solution structure predict values of 1.399x10(7) s-1, 1.500x10(7) s-1 and 1.863x10(7) s-1 when a solvent shell of 3.5 A is included in the calculation. The optimal orientation of the diffusion tensor has been incorporated into a hybrid Lipari-Szabo type local motion-anisotropic rotational diffusion model to characterize the local mobility in the molecule. The mobility parameters thus extracted show a quantitative improvement with respect to the model-free analysis assuming isotropic reorientation; helical regions exhibit similar dynamic properties and fewer residues require more complex models of internal motion. While the molecule is essentially rigid, a tripeptide loop region (residues 101 to 103) exhibits flexibility in the range of 20 to 30 ps, which appears to be correlated with the order in the NMR solution structure.

Tyrosine 64 of cytochrome c553 is required for electron exchange with formate dehydrogenase in Desulfovibrio vulgaris Hildenborough

Replacement of tyrosine 64 by alanine in cytochrome c553 from Desulfovibrio vulgarisHildenborough prevents electron transfer with the formate dehydrogenase. Biophysical and biochemical studies show that the protein is correctly folded and that the oxidoreduction potential is not modified. The solution structure of the mutant cytochrome determined by two-dimensional (2D) NMR clearly establishes that the overall fold of the molecule is nearly identical to that of the wild-type cytochrome. The electrostatic surface charge distributions for the wild-type and mutant cytochrome are similar, suggesting that the interaction site of the physiological partners is not modified by the mutation. The lack of the aromatic ring induces slight destabilization of the hydrophobic core of the molecule and modifications of the hydrogen bond at position 64, as well as conformational disorder of the side chain of K63. The loss of the hydrogen bond from tyrosine 64 and the increase of the solvent exposure of the heme are probably responsible of the loss of electron transfer between formate dehydrogenase and cytochrome c553.

NMR structures of ferredoxin chloroplastic transit peptide from Chlamydomonas reinhardtii promoted by trifluoroethanol in aqueous solution

The 32-amino acid transit peptide of the unicellular green alga Chlamydomonas reinhardtii ferredoxin has been synthesized and analysed by NMR spectroscopy and circular dichroism. The results show that while the peptide is unstructured in water, it undergoes an alpha-helix formation from residue 3 to 13 in a 30:70 molar-ratio mixture of 2,2,2-trifluoroethanol. The remainder of the peptide is still unstructured in CF3CD2OD/H2O mixtures, but is distributed on a side opposite to a hydrophobic ridge formed by Met5, Phe9 and Val13 on the induced alpha-helix. The NMR structures driven by 2,2,2-trifluoroethanol in aqueous solution, are discussed in terms of potent interactions with the chloroplast envelope and its translocation molecular machinery.