{"paper_id":"80e59120-b08d-4c21-b804-b99df80cc78d","body_text":"Visionerves: Automatic and Reproducible\nHybrid AI for Peripheral Nervous System\nRecognition Applied to Endometriosis Cases\nGiammarco La Barbera1,2(\f), Enzo Bonnot3,2,1, Thomas Isla1,\nJuan Pablo de la Plata1,2, Joy-Rose Dunoyer de Segonzac1, Jennifer Attali4,\nCécile Lozach4, Alexandre Bellucci5, Louis Marcellin6,7, Laure Fournier5,\nSabine Sarnacki8,1,2, Pietro Gori3,2, and Isabelle Bloch9,3,1,2\n1 IMAG2, Institut Imagine, Université Paris Cité, France\n2 Replico SAS, Paris, France\n3 LTCI, Télécom Paris, Institut Polytechnique de Paris, France\n4 Université Paris Cité, Department of Pediatric Imaging, Hôpital Necker\nEnfants-Malades, Assistance Publique-Hôpitaux de Paris (AP-HP), France\n5 Université Paris Cité, Department of Radiology, Hôpital Européen Georges\nPompidou, AP-HP, France\n6 Université Paris Cité, Department of Gynecological Surgery and Oncology\n(Professor Chapron), Hôpital Cochin, AP-HP, France\n7 Department of Development, Reproduction and Cancer (Professor Batteux),\nInstitut Cochin, Paris, France\n8 Université Paris Cité, Department of Pediatric Surgery, Hôpital Necker\nEnfants-Malades, AP-HP, France\n9 Sorbonne Université, CNRS, LIP6, Paris\ngiammarco.labarbera@replico.tech\nAbstract.Endometriosis often leads to chronic pelvic pain and possible\nnerveinvolvement,yetimagingtheperipheralnervesremainsachallenge.\nWe introduce Visionerves, a novel hybrid AI framework for peripheral\nnervous system recognition from multi-gradient DWI and morpholog-\nical MRI data. Unlike conventional tractography, Visionerves encodes\nanatomical knowledge through fuzzy spatial relationships, removing the\nneed for selection of manual ROIs. The pipeline comprises two phases:\n(A) automatic segmentation of anatomical structures using a deep learn-\ning model, and (B) tractography and nerve recognition by symbolic spa-\ntial reasoning. Applied to the lumbosacral plexus in 10 women with\n(confirmed or suspected) endometriosis, Visionerves demonstrated sub-\nstantial improvements over standard tractography, with Dice score im-\nprovements of up to 25% and spatial errors reduced to less than 5 mm.\nThis automatic and reproducible approach enables detailed nerve analy-\nsis and paves the way for non-invasive diagnosis of endometriosis-related\nneuropathy, as well as other conditions with nerve involvement.\nKeywords:Nerves Recognition·Hybrid AI·DWI·MRI·Endometriosis\narXiv:2509.18185v1  [cs.CV]  18 Sep 2025\n\n2 G. La Barbera et al.\n1 Introduction\nEndometriosis is a prevalent gynecological disease characterized by endometrial\ntissue outside the uterus, leading to chronic inflammation, immune dysfunction,\nand potential neurological involvement in the pelvic region. These neurological\nfactors are difficult to diagnose with conventional imaging. Moreover, a better\nunderstanding of how pelvic nerve fibers are affected in endometriosis could help\nexplaining the role they play in chronic pelvic pain [5,7]. Advanced techniques\nsuch as Diffusion Weighted MRI (DWI) and tractography [22] enable visualiza-\ntion of nerve fibers, but their application has mainly been limited to the central\nnervous system (CNS) [9,20,24], with limited studies in the peripheral nervous\nsystem (PNS) [4,19] and thus on endometriosis [18,26]. This is due to: (i) the\nchallenge of accurately and reproducibly reconstructing PNS via tractography\nalgorithms, primarily due to reliance on manual placement of regions of interest\n(ROIs) and significant inter-subject variability; (ii) the complexity of recognizing\neach individual nerve bundle, due to the abundance of streamlines and spurious\nfibers (such as muscle, tissue and noise) [12,17].\nIn this paper, we introduce “Visionerves\", an original hybrid AI method that\nleverages anatomical knowledge for automatic nerve identification, eliminating\nthe need for manual ROIs. By describing nerve trajectories relatively to other\nanatomical structures (segmented via deep learning from a standard MRI image)\nand modeling anatomical imprecision with fuzzy logic [3,13], our framework\nallowsguidingthetractographyalgorithm,makingtheresultsmorereproducible,\nand filtering out spurious (i.e. outliers) fibers by recognizing only nerve ones, via\nsymbolic spatial reasoning. Furthermore, using a software such as 3DSlicer [8],\nVisionerves results (anatomical segmentation and nerve fibers) can be merged\nand rendered into 3D models, offering valuable insights into the relationships\nbetween lesions, surrounding organs, and nerve structures [23]. This approach\ncould help explain the neuropathic component of chronic pain and paves the\nway for significant improvements in surgical and clinical planning. The presented\nmethod, applied in the pelvic region (from L5 to S3 nerve fibers) to 10 female\nadult subjects affected by endometriosis, yielded promising results, showing a\ngood correlation with nerve reference reconstructions.\n2 Related work\nTraditional methods employ the virtual dissection technique which requires man-\nual placement of ROIs to select or exclude fibers [12,17,18], but this process is\nlabor-intensive, time-consuming, and difficult to reproduce for complex tracts.\nAtlas-based approaches, commonly used in the brain, transfer labels via non-\nlinear deformations and mappings [21], but the accuracy of the results depends\nonalignmentandclusteringquality,whichisproblematicinpathologicalcases[9,\n10]. Machine learning and deep learning methods, explored for the CNS [16,25],\nrequire large annotated datasets, are often hard to interpret, and can be affected\nbydatabiasessuchassite-specificprotocolorscannerdifferences.Moreover,they\n\nVisionerves 3\ndo not address the inherent vagueness of PNS tract definitions, where boundaries\nare inherently ambiguous and difficult to delineate.\nTaking a different approach, WMQL [24] introduced a query language for\ndefining brain white matter tracts using mathematical models of spatial rela-\ntionships and logical operations. However, the method is limited to binary rela-\ntions, rough representation of structures, and does not account for pathological\ndeviations. A fuzzy set theory extension [6] addresses the inherent vagueness of\nanatomical definitions, but remains tailored to CNS fibers, not accounting for\nthe more complex spatial relationships seen in the PNS. For these reasons, these\nmethods offer valuable insights, but are insufficient to reconstruct and recognize\nconvoluted and smaller nerves of the PNS, such as the nerves of the pelvic lum-\nbosacral plexus, which are clinically significant in conditions like endometriosis.\n  \nA\nNERVE FIBER QUERIES\nL5_left = crossing(VertebralCanalL5Left)\nthen anterior of(PiriformisMuscleLeft)\nthen left of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n…\nL5_right = crossing(VertebralCanalL5Right)\nthen anterior of(PiriformisMuscleRight)\nthen right of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS1_left = crossing(SacralHoleS1Left\nthen anterior of(PiriformisMuscleLeft)\nthen left of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS1_right = crossing(SacralHoleS1Right)\nthen anterior of(PiriformisMuscleRight)\nthen right of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS2_left = crossing(SacralHoleS2Left)\nthen anterior of(PiriformisMuscleLeft)\nthen left of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS2_right = crossing(SacralHoleS2Right)\nthen anterior of(PiriformisMuscleRight)\nthen right of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS3_left = crossing(SacralHoleS3Left)\nthen anterior of(PiriformisMuscleLeft)\nthen left of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nS3_right = crossing(SacralHoleS3Right)\nthen anterior of(PiriformisMuscleRight)\nthen right of(LevatorAniMuscles)\nthen not posterior of(Sacrum)\n...\nB1 B2\nB1\nFig. 1.The Visionerves framework (blue) from image acquisition on the left to the final\n3D model merging anatomical structures and nerve fibers on the right. The method is\ncomposed of: phase A (green) - Anatomical Structure Segmentation - where a U-Net-\nbased algorithm takes as input a morphological (e.g. T2-w) MR image and outputs\nan anatomical segmentation; phase B (red) - Peripheral Nerves Reconstruction (B1)\nand Recognition (B2) - where a tractography algorithm takes as input a DW image\n(together with output segmentation of phase A), and outputs a fibers reconstruction,\nwhich in turn is passed as input (together with nerve fibers queries and result of phase\nA) to our symbolic AI filtering method for recognizing the targeted nerves.\n\n4 G. La Barbera et al.\n3 The Visionerves Method\nThe proposed Visionerves method consists of two primary phases (Figure 1): (A)\nanatomical structure segmentation from a morphological (e.g. T2-w) MR image\n(see Section 3.1); and (B) peripheral nerves reconstruction and recognition from\nmulti-gradient DW image, leveraging the result of phase A (see Section 3.2). As\ndescribed before, a final 3D model can be created, merging the extracted nerve\ninformation with the anatomical segmentation.\n3.1 Phase A - Anatomical Structure Segmentation\nAU-Net-basedalgorithminspiredbythewell-knownnnU-Net[14]anditsderiva-\ntives (nnDetection [1] for localization and nnInteractive [15] for error correction)\nis employed for rapid and fully automated modeling of anatomical structures\nfrom morphological MRI images, such as T2-w or T1-w. A preprocessing step\n(to have a standard dataset representation as in [14]) includes a reorientation\nof the images in the RAS coordinate system (left to right, posterior to ante-\nrior, inferior to superior), a correction of field heterogeneity, and a resampling\nto a common voxel size. The segmentation pipeline includes three custom U-Net\nsubsystems: one for localization (bounding box determination), one for semantic\nsegmentation, and one for error correction via user interaction.\nThis phase A of the pipeline enables fast and user-friendly 3D anatomical\nmodel generation, avoiding manual segmentation, however it is not central to our\nmethodanditcanbeeasilyreplacedwithalternativestate-of-the-artapproaches.\nFor these reasons we will not elaborate on it further. Nevertheless it is important\nto mention that in addition to enabling the functioning of phase B, this part also\nallows for the visualization of the nerves in their anatomical context, leading to a\nbetter understanding of their relationships to the different anatomical structures.\n3.2 Phase B - Peripheral Nerves Reconstruction and Recognition\nFor the identification of the PNS, in contrast to previous works [20,24], we pro-\npose to represent and formalize anatomical knowledge, usually given in natural\nlanguage, in first order logic, with the associated syntactic reasoning abilities.\nTaking inspiration from [6], we also propose to associate it with fuzzy semantics.\nSpatial relations between nerves and anatomical structures play an important\nrole in the description of nerves. They are represented as predicates in the logic,\nfor which a degree of satisfaction is computed using mathematical morphology\nand fuzzy sets [3]. These relations include distances, directions and connectiv-\nity with respect to segmented structures from phase A. This hybrid approach\ncombines formal descriptions (syntactic part) with concrete representations in\nthe spatial domain and as degrees of satisfaction taking values in[0,1](semantic\npart). Fuzzy representations hence inherently solve the semantic gap, establish-\ning links between abstract clinical concepts and image information.\n\nVisionerves 5\nPhase B1 - Automatic T ractography for Reconstruction.The first part\nof phase B consists of fiber reconstruction via a tractography algorithm [22] ap-\nplied to the multi-gradient DWI image. A preprocessing (as in [22]) includes\ndenoising, Gibbs artifact removal, eddy currents and motion corrections, and\ncorrection of field heterogeneity. Following this, a tractography algorithm is ap-\nplied (algorithm choice left to the user depending on DWI parameters and the\nstudied body region). In order to narrow the reconstruction space as well as make\nthis process automatic and reproducible, ROIs for seeding and inclusion zones\n(which are obligatory to be crossed in the specified order) are produced using\neither directly the segmented anatomical structures or by using spatial relations\nto define regions where they are satisfied (e.g. region “anterior of structure A”\nAND “to the right of structure B”). In this part, spatial relations are binarized, in\norder to be used in tractography algorithms. In addition, the created regions are\nlarge enough to not limit the reconstruction space too much, given the potential\ndeviations from normal anatomy. It is important to note that, in order to have\na good voxel matching between anatomical structures (produced from the MRI)\nand DWI, a registration might be necessary.\nPhase B2 - Symbolic AI for Recognition.Once the tractogram is recon-\nstructed, we perform a filtering to recognize only the nerve fibers, differentiat-\ning them from muscle and tissue fibers as well as potential noise. The medical\nknowledge encoded in the logic is translated into queries used by the recognition\nalgorithm. For each nerve bundle, a query representing its anatomical path is\ncreated using the spatial relations, with the possibility of combining the relations\nwith AND/OR operators, creating exclusion zones with NOT operators, and or-\ndering the nerve segments with THEN (i.e. “sequential\" AND) operators. Once\nthe query is built, its degree of satisfaction is assessed at each point along the\nfiber by mapping every point to the corresponding defined fuzzy regions. A fiber\nis considered to satisfy a query if it sequentially validates all specified spatial\nrelations, where to validate means that the average of the non-zero fuzzy values\nalong the fiber is higher than a specified threshold. All fibers that fulfill these\ncriteria are then aggregated to form a bundle representing the targeted nerve.\nMore details on the fuzzy logic modeling and on phase B2 can be found in [2].\n4 Application on the pelvic region for endometriosis\nWe applied our Visionerves method on the lumbosacral plexus in the 10 en-\ndometriosis cases, for whom a nerve fiber analysis could be a strong aid in the\nunderstanding of this disease as explained in Section 1. Furthermore, since this\nis a highly innervated region with different muscles and organs, it represents an\nideal subject of application for our method.\n4.1 Database\nFor the segmentation system of phase A, we used 168 T2-w MRI images of 131\npatients (ranging from 2 months old to 20 years old) belonging to a proprietary\n\n6 G. La Barbera et al.\ndatabase licensed by the Hôpital Necker-Enfants malades of Paris), with refer-\nence images manually annotated by expert surgeons and radiologists over the\ncourse of several years using 3DSlicer software [8]. Pelvic bones (L5 vertebra, hip\nbones and sacrum), muscles (piriformis, obturator and levator ani), visceral or-\ngans (bladder, colon and rectum) and reproductive organs (ovaries, uterus and\nvagina) were labeled, in addition to other regions of interest (sacral foramina\nfrom S1 to S3, sacral canal, intervertebral foramina of L5) facilitating nerve\ndetection.\nWe then applied the complete Visionerves method (phases A and B) on\n10 different adult female patients (5 diagnosed and 5 suspected endometriosis,\nboth groups in an age range from 20 to 50 years old), gathered at Hôpital\neuropéen Georges-Pompidou of Paris. For each patient, a couple of T2-w MRI\n(reconstruction voxel size 0.5×0.47×0.47mm3) and multi-gradient DWI image\n(acquisition voxel size 3.3×2.3×3.6mm3, NEX 1, 50 directions, b-value 600) was\nacquired in a 3T GE Signa Architect machine during a preliminary research (for\na future prospective study) and used retrospectively after anonymization. Nerve\nreference reconstructions were created using the method of phase B1 plus the\nuse of a ROI mask, in order to select the fibers passing through, exclusively and\ncompletely, within it. This further constrains the search area to the region we\nconsider to represent the true pathway of fiber passage. Such a ROI mask was\nproduced under the supervision of expert surgeons and radiologists via manual\nsegmentation of tubes enclosing each bundle of nerve fibers (given the difficulty\nof accurately segmenting these structures). All the manual annotations detailed\nin this paragraph were performed using 3DSlicer software [8].\nFinally the nerve queries were written with the help of clinical experts, lever-\naging anatomy books [11], literature [18] and knowledge, and aiming to make\nthem generalizable across different cases. For example a query for recognizing\nthe left S2 nerve is (the parameters defining the relations and threshold values\nare not mentioned here for the sake of readability):\nS2_left=crossing(SacralHoleS2Left) then anterior_of(PiriformisMuscleLeft)\nthen left_of(LevatorAniMuscles) then not posterior_of(Sacrum)\nthen not (crossing(SacralHoleS1Left) or crossing(SacralHoleS3Left))\nthen not left_of(PiriformisMuscleLeft)\nthen not anterior_of(ObturatorMuscleLeft)\nthen not between(ObturatorMuscleLeft, ObturatorMuscleRight)\n4.2 Results and Discussion\nThe networks in phase A were implemented from scratch using Tensorflow 2.16.\nWe used 89 T2-w images for training, 16 as validation set and 63 as test set.\nAll images were preprocessed as described in Section 3.1 with a common voxel\nsize of 0.88×0.88×0.88mm3. The method showed high quality segmentation of\nthe pelvic structures described in Section 4.1 with Dice indices exceeding 85%\nfor dense structures and Average Surface Distance less than 2 mm for elon-\ngated or small structures. The pelvic region was consistently localized by the\n\nVisionerves 7\nfirst network and the U-Net-based error correction proved effective for minor\npelvic structures, providing satisfying results for clinicians in around 2 minutes\nin worst case scenarios. Although the error correction phase reduces the level of\nautomation of phase A, it serves to decouple potential segmentation errors from\ndirectly propagating into the results of phase B.\nT able 1.Quantitative results of phase A of the Visionerves method in average (and\nstandard deviation) for 10 endometriosis cases on different anatomical structures using\nDice and Average Symmetric Surface Distance (ASSD). Results are shown before the\nuse of the U-Net-based error correction. Bones are L5 vertebra, hip bones and sacrum;\nmuscles are piriformis, obturator and levator ani; visceral organs are bladder, colon and\nrectum; reproductive organs are ovaries, uterus and vagina; specific ROIs are sacral\nforamina from S1 to S3, sacral canal and intervertebral foramina of L5.\nStructure Bones Muscles Visceral organsReproductive organsSpecific ROIs\nDice [%] 95.7 91.3 89.1 86.4 86.5\n(0.14) (0.54) (0.92) (0.78) (1.26)\nASSD [mm] 0.22 0.36 0.97 0.93 0.43\n(0.09) (0.42) (1.18) (0.62) (0.51)\nThe results of the complete Visionerves method on the 10 subjects with both\nT2-w and DWI acquisitions are shown in Table 1 for phase A (segmentation).\nThese results are reported for completeness, even though segmentation is not\nthe central focus of our method; they are shown before error correction, which\nwas rarely necessary. Phase B results (reconstruction and recognition) are shown\nin Table 2 and were obtained using the 10 nerve reference reconstructions avail-\nable from our patient cohort. After preprocessing, nerve bundles were extracted\nusing raw tractography. The sacral or intervertebral foramen corresponding to\neach nerve (for each side) served as the seed labelmap, and fiber selection was\nconstrained to those containing at least one point within the region traversed\nby the sciatic nerve (where all the four analyzed fiber bundles are confluent, see\nFigure 2 for a better understanding). This region was constructed using the bi-\nnarized spatial relations that we defined in Section 3.2. This phase was executed\nusing MRTrix3 software [22] and we used a FOD-based algorithm with deter-\nministic tracking, called “SD STREAM\" [21], with minimum FOD amplitude\nfor seeds of 0.15, FOD cut-off of 0.10, maximum angle of 45 degrees and step\nsize of 3mm. This raw tractogram is referred as just “Tractography” in Table\n2, and represents the current state of the art in PNS nerve recognition for tra-\nditional methods. Learning-based approaches could not be evaluated, primarily\ndue to the challenges associated with detailed manual nerve segmentation in the\nT2-w images, and, more critically, the infeasibility of training a neural network\ngiven the only 10 cases with DWI and the lack of publicly available pre-trained\nmodels. We then applied our filtering method based on symbolic AI using the\nnerve queries defined as in Section 4.1 (see also Figure 1) to the raw tractograms\nproduced. These results are referred as “+ Filtering” in Table 2. Since the Dice\n\n8 G. La Barbera et al.\nscore is not well adapted to the thin and tubular structure of the nerves, we\nalso used the precision score (recall was not considered, as it cannot exceed the\nperformance achieved by tractography) and multiple distance metrics that are\nthe Average Symmetric Surface Distance (ASSD), the Average Symmetric Cen-\nterline Distance (ASCD) and the Absolute Length Difference (ALD). In order\nto make these measurements, each fiber bundle was transformed into a single\nlabelmap.\nT able 2.Quantitative results of the Visionerves method (pre- and post-filtering) in\naverage (and standard deviation) for 10 endometriosis cases on 4 different lumbosacral\nnerves (divided by sides) using Dice, precision, Average Symmetric Surface Distance\n(ASSD), Average Symmetric Centerline Distance (ASCD) and Absolute (Euclidean)\nLength Difference (ALD).\nNerve\nBundle\nVisionerves Dice [%] Precision\n[%]\nASSD\n[mm]\nASCD\n[mm]\nALD [mm]\nL5\nleft\nTractography 49.56\n(19.22)\n35.09\n(18.69)\n9.66\n(4.42)\n31.67\n(19.33)\n30.64\n(33.73)\n+ Filtering 70.70\n(16.26)\n63.65\n(17.35)\n5.02\n(4.38)\n14.27\n(16.11)\n26.49\n(30.52)\nL5\nright\nTractography 55.81\n(14.29)\n40.21\n(14.21)\n6.30\n(3.63)\n18.28\n(12.53)\n22.25\n(27.78)\n+ Filtering 64.86\n(9.85)\n58.74\n(14.31)\n3.49\n(2.38)\n10.01\n(8.89)\n24.98\n(30.65)\nS1\nleft\nTractography 56.37\n(21.52)\n42.37\n(21.95)\n6.02\n(3.48)\n14.14\n(8.34)\n36.12\n(22.75)\n+ Filtering 74.31\n(11.95)\n70.02\n(16.47)\n2.37\n(1.80)\n7.20\n(7.00)\n21.37\n(23.47)\nS1\nright\nTractography 60.40\n(24.97)\n47.79\n(27.36)\n6.59\n(4.39)\n13.44\n(11.44)\n14.98\n(18.03)\n+ Filtering 74.46\n(19.94)\n69.65\n(25.02)\n2.12\n(2.54)\n5.81\n(5.32)\n15.03\n(18.34)\nS2\nleft\nTractography 50.96\n(18.43)\n36.57\n(20.23)\n7.27\n(3.11)\n19.30\n(15.20)\n35.76\n(35.40)\n+ Filtering 69.32\n(12.48)\n59.45\n(17.39)\n3.52\n(2.27)\n11.95\n(18.01)\n32.40\n(27.03)\nS2\nright\nTractography 56.78\n(17.69)\n41.83\n(17.67)\n6.37\n(4.11)\n13.54\n(8.82)\n25.30\n(17.36)\n+ Filtering 70.71\n(10.05)\n61.19\n(12.66)\n3.01\n(2.61)\n6.37\n(5.75)\n17.77\n(17.68)\nS3\nleft\nTractography 44.54\n(18.73)\n30.55\n(16.40)\n7.48\n(3.32)\n17.47\n(7.07)\n37.68\n(32.23)\n+ Filtering 54.78\n(15.95)\n56.90\n(21.35)\n3.19\n(1.83)\n7.14\n(2.82)\n30.61\n(21.72)\nS3\nright\nTractography 39.49\n(18.56)\n26.32\n(15.11)\n9.66\n(4.40)\n21.69\n(12.07)\n33.70\n(27.81)\n+ Filtering 54.06\n(18.77)\n55.26\n(23.43)\n4.28\n(3.54)\n6.88\n(2.67)\n28.55\n(22.80)\n\nVisionerves 9\nResults show that just using tractography, even guided via ROIs for seeding\nand inclusion, is not enough to eliminate all spurious fibers. By contrast, the\nproposed approach via symbolic AI allows Dice results to be increased by at\nleast 15%and precision by at least 25%for most fibers, indicating proper elim-\nination of false positives with nearly no loss of true positives. Distance metrics\n(more suitable for evaluation of elongated structures) decrease drastically to half\na centimeter for the entire bundle surface and a centimeter when considering the\ncenterline. The difference in length remains around 2 cm, indicating that some\nfibers are not tracked all the way to the end of the portion of the fiber associated\nwith the sciatic nerve. According to clinical experts, these distances are accept-\nable (to a certain extent) for clinic use, in particular when examined along with\na visual assessment. Notably, Dice and precision scores are lower for S3 due to its\nsmaller fiber diameter, making these evaluation measures less suitable; however,\ndistance-based measures remain comparable to those of other nerve bundles.\n  \nTractography + Filtering Reference\nBEST CASES\nW\nORST CASES\nFig. 2.Qualitative results of the Visionerves method (before and after recognition\nfiltering) for the two best cases and two worst cases of the 10 endometriosis cases.\nNerve reference reconstructions are produced as described in Section 4.1. Color code\nfor nerve bundle: L5 in green, S1 in blue, S2 in magenta and S3 in cyan.\nSome qualitative results are presented in Figure 2: for the two best cases,\nexcept for a few spurious fibers, the bundles overlap almost perfectly, while in\nthe two worst cases most false positives are still effectively filtered (moreover\non one case reference nerves are also highly disorganized). Only one case (third\nrow) appears anatomically incorrect, but the reference nerves are also very dis-\norganized. Interestingly, the two best-performing cases correspond to patients\nwith confirmed endometriosis, whereas the worst-performing cases involve pa-\ntients with suspected endometriosis (who exhibited symptoms of menorrhagia\nand pelvic pain). It remains unclear whether these suboptimal results are due to\n\n10 G. La Barbera et al.\nthe choice of the tractography algorithm, its parameter settings or to intrinsic\nlimitations of the imaging data. A more in-depth study needs to be conducted.\nMoreover, the filtering parameters and threshold values may probably need to be\nadjusted in cases where the raw tractogram already appears very disorganized.\n5 Conclusion and Perspectives\nIn this work, we introduced Visionerves, a novel hybrid AI framework that lever-\nages anatomical knowledge for the automatic and reproducible recognition of\nperipheral nerves. By integrating fuzzy spatial reasoning with symbolic AI and\nmulti-modal MRI, our method addresses key limitations of traditional tractogra-\nphy, most notably its reliance on manual intervention and lack of reproducibil-\nity. For a preliminary assessment we applied Visionerves in the pelvic region\non 10 endometriosis cases. Results demonstrated significant improvements over\nconventional tractography approaches confirming that our method can substan-\ntially reduce spurious fibers while maintaining the structural integrity of the\nnerve reconstructions.\nBy enabling the reproducible and individualized extraction of nerve bundles,\nVisionerves paves the way for assessing nerve-specific diffusion and morpholog-\nical characteristics. This is particularly valuable in endometriosis, where these\nfiber properties may enable diagnosis in suspected cases of nerve involvement\nwithout the need for surgical biopsy. The S2 and S3 would be the principal\nnerves studied for this pathology since they are part of the pudendal plexus that\nis innervating the external genitalia and the perineum. Even though L5 and S1\nfurther constitute the sciatic nerve, they are still of interest due to their proxim-\nity to the uterus and ovaries. Other conditions with possible nerve involvement\nmay also benefit from our method (e.g. pelvic cancers), where Visionerves can\nalso be used for pre-surgical planning and longitudinal follow-up.\nIn order to reach clinically acceptable results, in future work we plan to:\n(i) extend the database, including also healthy subjects and other pathologies,\nin order to enable more informed selection of hyperparameters during both phase\nB stages; (ii) refining query formalization (e.g. new spatial relations), and auto-\nmate its generation from textual anatomical descriptions and its parametrization\nin cases where the raw tractogram is full of spurious fibers; (iii) exploring alter-\nnative validation criteria for fiber selection or adopt a solution where threshold\nvalues are adjusted based on a general “anatomical coherence score” that can be\nadjusted by the user, such as in [6]. Ultimately, we will test Visionerves on other\npelvic nerve fibers (e.g. S4, pudendal, obturator) and other regions, (e.g. skull\nbase, head and neck, brachial plexus), where conventional tractography remains\nparticularly challenging. Once an anatomical segmentation of a region and the\ntrajectories of the nerves in that region are defined, the Visionerves method can\nbe applied in a straightforward manner, leveraging the existing comprehensive\nset of spatial relations, which can be easily expanded if needed.\nAcknowledgments.ThisworkhasbeenfundedandsupportedbyLiguecontrelecan-\ncer, Fondation Béatrice Denys and Prématuration IP Paris. This work was performed\n\nVisionerves 11\nusing HPC Jean-Zay resources from GENCI–IDRIS (Grant 2025-AD011015418). We\nwould like to thank Fatiha Tacine for her assistance with the organization and retrieval\nof the acquisitions at the Hôpital européen Georges Pompidou. We would like to thank\nalso Alice Sorrentino for her assistance with the organization and segmentation of the\nacquisitions at the Hôpital Necker-Enfants malades.\nData Use Declaration.The 131 patients used in phase A were included under a\nlicense granted by the Hôpital Necker-Enfants malades for acquisitions during proto-\ncol n°2015-101705-44. The 10 endometriosis patients used for testing the Visionerves\nframework were included under preliminary acquisitions of a future research clinical\nprotocol n°2024-100538-39 approved by the Hôpital européen Georges-Pompidou.\nDisclosure of Interests.The authors declare that a patent application has been filed,\nrelated to the method presented in this paper. This disclosure is made in the interest\nof transparency and does not affect the integrity or objectivity of the research.\nReferences\n1. Baumgartner, M., Jäger, P.F., Isensee, F., Maier-Hein, K.H.: nnDetection: A self-\nconfiguring method for medical object detection. In: Medical Image Computing\nand Computer Assisted Intervention. vol. 12905, pp. 522–532 (2021)\n2. Bloch, I., Bonnot, E., Gori, P., La Barbera, G., Sarnacki, S.: First order logic with\nfuzzy semantics for describing and recognizing nerves in medical images. In: IEEE\nInternational Conference on Fuzzy Systems (FUZZ-IEEE) (2025)\n3. Bloch,I.,Ralescu,A.:FuzzySetsMethodsinImageProcessingandUnderstanding:\nMedical Imaging Applications. Springer Cham (2023)\n4. Cage, T., Yuh, E., Hou, S., Birk, H., Simon, N., Noss, R., et al.: Visualization of\nnerve fibers and their relationship to peripheral nerve tumors by diffusion tensor\nimaging. Neurosurgical Focus39(3), E16 (2015)\n5. Chapron, C., Marcellin, L., Borghese, B., Santulli, P.: Rethinking mechanisms, di-\nagnosis and management of endometriosis. Nature Reviews Endocrinology15(11),\n666–682 (2019)\n6. Delmonte, A., Mercier, C., Pallud, J., Bloch, I., Gori, P.: White matter multi-\nresolution segmentation using fuzzy set theory. In: IEEE International Symposium\non Biomedical Imaging (2019)\n7. Fauconnier, A., Chapron, C.: Endometriosis and pelvic pain: epidemiological ev-\nidence of the relationship and implications. Human Reproduction Update11(6),\n595–606 (2005)\n8. Fedorov, A., Beichel, R., Kalpathy-Cramer, J., Finet, J., Fillion-Robin, J.C., Pu-\njol, S., Bauer, C., Jennings, D., Fennessy, F., Sonka, M., Buatti, J., Aylward, S.,\nMiller, J.V., Pieper, S., Kikinis, R.: 3D Slicer as an image computing platform for\nthe quantitative imaging network. Magnetic Resonance Imaging30(9), 1323–1341\n(2012)\n9. Garyfallidis, E., Côté, M.A., Rheault, F., Sidhu, J., Hau, J., Petit, L., Fortin, D.,\nCunanne, S., Descoteaux, M.: Recognition of white matter bundles using local\nand global streamline-based registration and clustering. NeuroImage170, 283–295\n(2018)\n10. Gori, P., Durrleman, S., Colliot, O., Mangin, J.F., Ayache, N.: A prototype repre-\nsentation to approximate white matter bundles with weighted currents. In: Medical\nImage Computing and Computer-Assisted Intervention. pp. 289–296 (2014)\n\n12 G. La Barbera et al.\n11. Gray, H.: Anatomy of the Human Body. Lea and Febiger (1918)\n12. Haakma, W., Dik, P., Ten Haken, B., Froeling, M., Nievelstein, R., Cuppen, I.,\nDe Jong, T., Leemans, A.: Diffusion tensor magnetic resonance imaging and fiber\ntractography of the sacral plexus in children with spina bifida. Journal of Urology\n192(3), 927–933 (2014)\n13. Hudelot, C., Atif, J., Bloch, I.: Fuzzy spatial relation ontology for image interpre-\ntation. Fuzzy Sets and Systems159(15), 1929–1951 (2008)\n14. Isensee, F., Jaeger, P., Kohl, S., Petersen, J., Maier-Hein, K.: nnU-Net: a self-\nconfiguringmethodfordeeplearning-basedbiomedicalimagesegmentation.Nature\nMethods18(2), 203–211 (2021)\n15. Isensee, F., Rokuss, M., Krämer, L., Dinkelacker, S., Ravindran, A., Stritzke, F.,\nHamm, B., Wald, T., Langenberg, M., Ulrich, C., Deissler, J., Floca, R., Maier-\nHein, K.: nnInteractive: Redefining 3D promptable segmentation. arXiv preprint\narXiv:2503.08373 (2025)\n16. Knoedler, M., Feibus, A., Lange, J., Venkatesh, R., Landman, J.: Individualized\nphysical 3-dimensional kidney tumor models constructed from 3-dimensional print-\ners result in improved trainee anatomic understanding. Urology85(6), 1257–1261\n(2015)\n17. Lemos, N., Melo, H., Sermer, C., et al.: Lumbosacral plexus MR tractography: A\nnovel diagnostic tool for extraspinal sciatica and pudendal neuralgia? Magnetic\nResonance Imaging83, 107–113 (2021)\n18. Manganaro, L., Porpora, M., Vinci, V., Bernardo, S., Lodise, P., Sollazzo, P., et al.:\nDiffusion tensor imaging and tractography to evaluate sacral nerve root abnor-\nmalities in endometriosis-related pain: A pilot study. European Radiology24(1),\n95–101 (2014)\n19. Muller, C., Mille, E., Virzì, A., Marret, J.B., Peyrot, Q., Delmonte, A., Berteloot,\nL., Gori, P., Blanc, T., Grevent, D., Boddaert, N., Bloch, I., Sarnacki, S.: Integrat-\ning tractography in pelvic surgery: a proof of concept. Journal of Pediatric Surgery\nCase Reports48, 101268 (2019)\n20. Smith, R., Tournier, J.D., Calamante, F., Connelly, A.: Anatomically-constrained\ntractography: improved diffusion MRI streamlines tractography through effective\nuse of anatomical information. NeuroImage62(3), 1924–1938 (2012)\n21. Tournier, J.D., Mori, S., Leemans, A.: Diffusion tensor imaging and beyond. Mag-\nnetic Resonance in Medicine65(6), 1532–1556 (2011)\n22. Tournier, J., Smith, R., Raffelt, D., et al.: MRtrix3: A fast, flexible and open\nsoftware framework for medical image processing and visualisation. NeuroImage\n202, 116–37 (2019)\n23. Vinit, N., Blanc, T., Bloch, I., Sarnacki, S., Agon, C., Romero, G.: Robotics and\n3D modeling for precision surgery in pediatric oncology. EJC Paediatric Oncology\n4, 100181 (2024)\n24. Wassermann, D., Makris, N., Rathi, Y., Shenton, M., Kikinis, R., Kubicki, M.,\nWestin, C.F.: The white matter query language: a novel approach for describing\nhuman white matter anatomy. Brain Structure and Function221(9), 4705–4721\n(2016)\n25. Wasserthal, J., Neher, P., Maier-Hein, K.: TractSeg - fast and accurate white mat-\nter tract segmentation. NeuroImage183, 239–253 (2018)\n26. Zijta, F., Lakeman, M., Froeling, M., van der Paardt, M., Borstlap, C., Bipat,\nS., et al.: Evaluation of the female pelvic floor in pelvic organ prolapse using 3.0-\nTesla diffusion tensor imaging and fibre tractography. European Radiology22(12),\n2806–2813 (2012)","source_license":"CC0","license_restricted":false}