A systematic approach to standardizing the visual appearance of endometriotic lesions for artificial intelligence recognition

N.B. and M.C. developed the original design. The acquisition, analysis and interpretation of the data were conducted by F.D. and A.N.A.N., F.D. and H.A. wrote the first draft of the report. All the authors contributed to the writing of the final manuscript.

This work was awarded two research grants: Gedeon Richter France and CNGOF Grant for Benign pathologies of the uterus and endometrium (2020), and AAP Innovation (public hospitals of the Marseille APHM). This work is part of and has been supported by the FEMaLe project (Finding Endometriosis using Machine Learning project) funded by The European Union's Horizon 2020 Research and Innovation Programme under Grant agreement 101017562.

The initial literature search revealed 932 articles (Figure  1 ). After an initial reading of the titles and abstracts, 42 articles were deemed relevant, and the manuscripts of 41 articles were read in full. The manuscript of one article, dating from 1989, could not be obtained. 17 Following this reading, 35 articles were included in this search (Table  1 ). Most articles relevant to the visual description of endometriosis via laparoscopy concern superficial endometriosis, with only a few articles concerning other visual appearances (microscopic endometriosis, adhesions, endometrioma, or deep endometriosis). PRISMA 2020 flow diagram. Studies included in the review. 635 patients. Laparoscopy for pelvic pain and infertility. 37 patients. Laparoscopy for infertility. 15 peritoneal biopsies. Laparoscopy for infertility. 137 biopsies, 77 patients. Laparoscopy for infertility, pelvic pain or tubal sterilization. 14 patients Laparoscopy for endometriosis resection 109 patients ‐ 164 biopsies. Laparoscopy with postoperative diagnosis of endometriosis. 33 patients. Laparoscopy for pelvic pain. 1440 patients. Laparoscopies performed by the same surgeon. 118 patients. Laparoscopy for infertility. 179 patients. Laparoscopy for infertility and/or pelvic pain. 153 patients. Laparoscopy for infertility or pelvic pain. 40 patients. Laparoscopy before and 6 months after cessation of medical treatment of endometriosis. 51 patients. Laparoscopy in patients with “chocolate cysts” Prospective, descriptive 100 patients ‐ 119 biopsies. Laparoscopy for infertility, pelvic pain or tubal sterilization. Number of patients not specified. Laparoscopy for pelvic pain or infertility. 62 patients ‐ 150 biopsies. Laparoscopy for pelvic pain. 44 patients. Laparoscopy for pelvic pain. 65 patients ‐ 189 biopsies. Laparoscopy for pelvic pain then laparoscopy 6 months later. 215 patients. Laparoscopy with postoperative diagnosis of endometriosis. 118 patients ‐ 311 biopsies. Laparoscopy for pelvic pain or infertility. 164 patients. Laparoscopy for suspected endometriosis. 54 patients, 122 biopsies. Laparoscopy for pelvic pain. 37 patients. Laparoscopy for pelvic pain 512 patients, 2005 biopsies. Laparoscopy for pelvic pain. 156 patients, 238 biopsies. Laparoscopy for pelvic pain. 133 patients, 611 biopsies. Laparoscopy for pelvic pain 142 patients. Laparoscopy for chronic pelvic pain. Superficial endometriosis is classically defined as infiltration of the peritoneum to a depth of less than 5 mm. Several authors have also proposed a limit of 2 mm or less, defining depths between 3 and 4 mm as intermediate infiltrations. 28 These superficial endometriotic lesions include a wide variety of visual manifestations that are often classified as typical (or black lesions) or atypical or nonpigmented (other lesion types). 21 , 27 , 38 The black superficial lesions, which are easily identifiable, are the most frequently and earliest described lesions in the literature (Table  2 ). The prevalence of these lesions in patients undergoing laparoscopy for infertility and/or pelvic pain is approximately 40% according to two studies. 26 , 38 These typical lesions have excellent diagnostic value for the presence of endometriosis when identified by a surgeon during laparoscopy; most studies have shown a PPV over 85% based on histological confirmation. 23 , 26 , 27 , 38 , 44 , 46 Histological confirmation rates of typical black lesions. Abbreviations: PPV, Positive Predictive Value. Histologically, typical black lesions are the result of old tissue bleeding followed by the retention of blood pigments. They therefore contain a combination of glands, stroma, and intraluminal debris surrounded by a fibromuscular matrix. 43 , 52 John Fallon et al. first described nonblack endometriotic lesions as colorless “amenorrheic” lesions in 1950. 53 In 1969, Karnaki published an age‐dependent appearance starting with an initial water blister. 54 By 1980, this was seen as common in adolescents. 55 In 1986, Jansen et al. and Russell reported that more than half of patients who underwent laparoscopic surgery for endometriosis had atypical nonpigmented lesions. 21 The different lesions described were as follows: white opacification of the peritoneum, 81% of which were histologically confirmed as endometriosis; “red flamelike” lesions also histologically confirmed in 81% of the biopsies; “glandular excrescences” (67% histologically confirmed); subovarian adhesions (50% histologically confirmed); “yellow–brown patches” (47% histologically confirmed); and circular peritoneal defects (45% histologically confirmed). In the decade following this 1986 description, these so‐called “atypical” or “subtle” lesions were widely reported by many authors, who used a wide variety of different terminologies to describe them (Table  3 ). Several authors have distinguished typical black lesions from other lesions; these lesions are therefore described as atypical, whereas others speak of pigmented, nonpigmented, and undefined lesions. 34 , 42 , 56 Finally, many authors, to varying degrees and nuances, have used qualifiers by color: “black,” “red,” “white,” and “subtle.” Each lesion color comprises numerous subtypes whose differentiation is very difficult to clearly appreciate through the articles. 22 , 32 , 37 , 40 , 49 , 52 , 57 The diagnostic value of identifying these atypical lesions by laparoscopic surgeons is highly variable, depending on the article and the type of lesion. It seems to be generally lower than that of typical black lesions (Table  4 ). However, the PPV for the identification of red lesions seems to be greater than that for other types of atypical lesions. 21 , 23 , 44 Terminologies used to describe the different types of atypical endometriosis lesions. Martin et al., 1989 [ 26 ] Stratton et al., 2002 [ 39 ] Shafik et al., 2000 [ 37 ] Vernon et al., 1986 [ 22 ] Stegmann et al., 2008 [ 49 ] Walter et al., 2001 [ 38 ] Moen et al., 1992 [ 29 ] Brosens et al., 1997 [ 36 ] Wiegerinck et al., 1993 [ 31 ] Brosens et al., 1997 [ 36 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Brosens et al., 1997 [ 36 ] Martin et al., 1989 [ 26 ] Stripling et al., 1988 [ 23 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Brosens et al., 1997 [ 36 ] Marchino et al., 2005 [ 46 ] Brosens et al., 1997 [ 36 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Brosens et al., 1997 [ 36 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Stratton et al., 2002 [ 39 ] Shafik et al., 2000 [ 37 ] Brosens et al., 1997 [ 36 ] Martin et al., 1989 [ 26 ] Stripling et al., 1988 [ 23 ] Wood et al., 2002 [ 40 ] Walter et al., 2001 [ 38 ] Moen et al., 1992 [ 29 ] Marchino et al., 2005 [ 46 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Walter et al., 2001 [ 38 ] Moen et al., 1992 [ 29 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Walter et al., 2001 [ 38 ] Stratton et al., 2002 [ 39 ] Moen et al., 1992 [ 29 ] Chatman et al., 1981 [ 18 ] Brosens et al., 1997 [ 36 ] Martin et al., 1989 [ 26 ] Marchino et al., 2005. [ 46 ] Martin et al., 1989 [ 26 ] Moen et al., 1992 [ 29 ] Wood et al., 2002 [ 40 ] Walter et al., 2001 [ 38 ] Shafik et al., 2000 [ 37 ] Marchino et al., 2005 [ 46 ] Jansen et al., 1986 [ 21 ] Nisolle et al., 1990 [ 27 ] Donnez et al., 1996 [ 35 ] Walter et al., 2001 [ 38 ] Brosens et al., 1997 [ 36 ] Positive predictive value (PPV) of visual diagnosis of atypical endometriosis lesions at laparoscopy. Polypoid: 75% (9/12) Flat: 33% (4/12) Raised: 33% (2/6) Non‐pigmented: 63%. Undefined" lesions: 52%. Atypical lesions each have a different visual appearance and histological characterization, which calls into question the value of grouping them in a common class. 18 , 27 , 43 , 52 , 58 , 59 Red lesions are thus active, proliferative endometriosis lesions consisting of glands and stroma with developed neovascularization and are often associated with recent intralesional bleeding. 19 , 27 , 43 , 58 On the other hand, white lesions are poorly vascularized, often inactive, often nonproliferative, and consist mainly of fibrosis and some degree of pigmentation. 58 The whitish opacity of the peritoneum consists of a retroperitoneal endometrial glandular structure associated with a poorly developed stroma surrounded by fibrotic tissue. 18 , 43 Sometimes, a hemosiderin blood pigment is present, which results in the appearance of yellow–brown or café‐au‐lait patches. The synthesis of data from the literature allowed us to establish a visual ontology of superficial endometriosis lesions, as shown in Figure  2 . The lesions are thus divided into five visual classes: microscopic (invisible), subtle, red, black, and white. Some classes include the main subtypes found in the literature (10 in all); an attempt has been made to group the synonyms assumed in the articles under a single subtype. A synthetic description of the histology of each subtype is also provided. Visual ontology for superficial endometriosis. In the context of endometriosis, adhesions are usually differentiated according to their density and transparency. 21 , 27 , 29 , 36 Therefore, adhesions can be dense or filmy, depending on their transparency, and some filmy adhesions can act as bridges for vessels. The 1985 revised American Fertility Society (AFS) classification takes up this distinction, granting a different number of points for the two types of adhesions. 60 Adhesions have no specific characteristics that would allow them to be differentiated from adhesions of infectious or surgical origin (which are also frequently observed in patients with endometriosis). Although they can occur at any location in the peritoneal cavity, subovarian adhesions are more specifically described for endometriosis by many authors. 6 They are thought to be the consequence of an inflammatory reaction induced by active lesions. 43 Because they are not very specific, the rate of histological confirmation of the endometriotic nature of adhesions is generally low, ranging from 16 to 50%, depending on the study (Table  5 ). Positive predictive value (PPV) of visual diagnosis of adhesions in the context of endometriosis. Very few articles have described the laparoscopic visual characteristics of endometriomas, which are classically recognized on imaging before surgery (by magnetic resonance imaging (MRI) or ultrasound). 61 In 1991, Vercellini et al. reported that endometriomas could be visually recognized laparoscopically with a set of 4 criteria: (i) had an ovarian cyst measuring less than 12 cm, (ii) had adhesions to the pelvic sidewall and/or broad ligament, (iii) had typical black superficial lesions on the surface, and (iv) had thick chocolate‐colored “tarry” contents. 62 According to the present study, these characteristics allow identification by surgeons in 97.5% of patients with histological confirmation. In a 1990 study of 41 ovarian cysts, Martin and Berry noted that five cysts (12%) were wrongly considered to be endometriomas during surgery, whereas on histology, they were corpus lutea or corpus albicans. 63 The authors gave their laparoscopic description of the endometrioma as follows: “flattened white internal lining with irregular raised red or red and brown streaks scattered throughout the internal wall”. 63 Other studies also reported excellent PPVs for the recognition of endometriomas by surgeons via laparoscopy (79.5 to 97.7% (Table  6 )). 30 , 33 , 39 , 48 , 62 Positive predictive value (PPV) of visual diagnosis of endometriomas at laparoscopy. < 2cm: 100%. 2–6cm: 0%. 6–12cm: 50%. 12–20cm: 85%. Typical Endometriomas: 89% Atypical endometriomas: 42%. Fewer than 10 studies have described the visual appearance of deep endometriosis during laparoscopy. 45 Usually, deep endometriosis is recognized primarily by clinical examination and palpation or by medical imaging (MRI or ultrasound), which is increasingly sensitive. 26 , 44 , 64 As early as 1979, the AFS classified the presence of dense adhesions obliterating the cul‐de‐sac as a criterion for the severity of endometriosis. 65 In 1990, Cornillie et al. reported that because of two‐dimensional vision and limited palpation sensitivity, the depth of deep endometriosis lesions is difficult to assess by laparoscopy. 28 According to these authors, retraction of the peritoneum at the surface is the only clue for identifying lesions. In a subsequent publication, Koninckx and Martin retrospectively analyzed 136 deep endometriotic lesions (histological slides and laparoscopy photographs). They concluded that there were three forms of deep endometriosis: (i) a conical form, suggesting an infiltration mechanism; (ii) a form covered with adhesions, suggesting a retraction mechanism; and (iii) a spherical form, most of which is buried under the peritoneum. The authors also noted that the volume of lesions of types (ii) and (iii) could not be properly appreciated visually from the peritoneal cavity since they were buried under adhesions and within the peritoneum itself, respectively. 66 Later, Donnez et al. described two additional procedures for identifying deep endometriotic lesions via laparoscopy: obliteration of the space and deformation of the organs. 35 , 67 The matter of visualizing deep endometriosis has also arisen more recently. Several authors have reported that certain deep endometriotic lesions, particularly rectal nodules, are identifiable only by palpation and cannot be visualized. 68 , 69 , 70 Roman et al. estimated that in approximately 25% of patients undergoing bowel resection, there are non‐visible but palpable satellite lesions measuring up to 1 cm that may be more than 2 cm away from the initial identified lesion. 10 The data cited above were correlated with our surgical video database to establish correspondences with the different types of lesions observed to verify that each lesion could be reasonably assigned to a specific class, with minimal ambiguity. This work made it possible to construct a visual ontology of endometriosis according to 4 visual classes, which were subdivided into 11 subclasses (Figure  3 ). Superficial endometriosis was further subdivided into 8 additional visual categories (considering that endometriosis on the normal peritoneum (invisible) is not a visual category and that the ‘black’ category does not contain any additional subcategory) (Figure  2 ). Visual ontology for endometriosis.

We have presented a systematic search of the visual descriptions of endometriotic lesions under laparoscopy. This article represents the methodological foundation of a broader AI‐based project whose next stage involved the automatic segmentation and classification of endometriotic lesions using deep learning. 13 Here, we focus specifically on the conceptual and visual framework required to ensure robust and reproducible human annotation—a critical yet often neglected prerequisite in AI model development. This structured ontology, grounded in surgical expertise, enables the precise definition of classes that guide image annotation and improve model interpretability. In this search, we propose a focused visual ontology with 4 main visual classes, which may serve as the basis for the automatic detection and recognition of endometriosis via AI. With respect to superficial endometriosis, our search of the literature revealed that many terms have been used to describe variations in color and shape. Although the depth of infiltration cannot be reliably assessed from the peritoneal cavity alone, and the designation of a lesion as “superficial” may therefore be uncertain, we chose to retain this category within the ontology. This decision was motivated by its widespread clinical use, its relevance for surgical decision‐making, and the need for the ontology to remain interpretable by a broad range of clinicians. We retained four main visual subclasses according to the three most commonly used colors (black, red, white), and the other lesions were grouped together as “subtle endometriosis.” Although the term subtle has already been used in the literature to describe any form of superficial endometriosis other than black, we have chosen to use it to describe lesions that do not correspond to any of the three dominant colors. We chose not to use the term atypical, as we felt that there were too many different visual aspects of endometriosis for any one form to be considered typical. It would appear that these forms are the least frequently histologically confirmed and can therefore legitimately be described as subtle. We then used findings from our video database to subdivide the main classes, obtaining a total of nine classes. It was clear from the literature and our videos that a distinction had to be made between dense and filmy adhesions. With respect to the description of deep endometriosis, 3 visual subclasses, which are most often found in the literature (deformation, retraction, obliteration), were distinguished. This differentiation is debatable because it appears obvious when reviewing our video database that these visual aspects are not mutually exclusive but are often associated to various degrees. Notably, deep endometriotic lesions are frequently centered on one or more superficial lesions, and this visual aspect is sometimes easier to detect. Finally, with respect to ovarian endometriosis, we have used the consensual term “endometrioma,” although it should be noted that this type of cyst is not always directly visible because it can be hidden beneath the ovarian cortex and can be confused with a hemorrhagic corpus luteum. Our ontology includes a wide range of lesion types, including subtle or nonpigmented forms whose clinical relevance remains uncertain. We consider this to be a faithful reflection of real‐life surgical experience. Furthermore, comprehensive lesion recognition is a prerequisite for any diagnostic application. This approach may also enable future studies to correlate lesion subtypes or volumes with symptoms or outcomes, thereby refining our understanding of their clinical significance. In a previously published companion paper, 13 this visual ontology was applied to develop an AI‐based system for automated recognition of endometriosis lesions during laparoscopy. To illustrate how the lesion‐specific predictive values reported in the literature relate to automated recognition, Table  7 provides a contextual comparison between the PPVs of surgeons' visual diagnosis and the precision achieved by the AI model trained using this ontology. Contextual comparison between visual diagnosis from surgeons (PPV from the literature) and AI‐based lesion recognition (precision). Abbreviations: AI, artificial intelligence; PPV, positive predictive values. The AI performance values reported in this table are extracted from a previously published study and are provided for contextual comparison only. AI precision was calculated using expert visual annotations as the reference standard. In the AI‐based study, adhesions were analyzed separately as dense and filmy, whereas the literature reviewed here does not allow reliable estimation of PPVs for these subtypes, and adhesions are therefore reported as a single combined category. One might reasonably expect that lesion types with the highest PPVs in the literature—reflecting higher histological specificity for endometriosis—would also be those most accurately recognized by AI. This assumption appears to hold to some extent for black superficial lesions, as well as for white and subtle lesions, for which AI precision values are broadly consistent with the ranges of human PPVs reported in the literature. In contrast, the markedly lower AI precision observed for red superficial lesions is unexpected given their relatively high PPVs in several surgical series. This discrepancy may reflect the limited visual specificity of red coloration in laparoscopy, where erythema, vascular structures, inflammation, or bleeding are ubiquitous and may confound automated recognition. A similar observation applies to ovarian endometriosis, where AI precision appears lower than human PPVs, potentially related to the explicit separation of endometriomas and chocolate fluid as distinct visual classes in the AI study, whereas these entities are often implicitly grouped in surgical practice and in the literature. Conversely, AI‐based recognition of adhesions demonstrated relatively high precision despite low PPVs reported in the literature based on systematic histological confirmation. This apparent contradiction highlights a fundamental difference in reference standards: while the literature evaluates visual diagnosis against a histological ground truth, the AI model was trained and evaluated using expert visual annotations as its reference standard. For lesions with poor histological specificity but consistent visual patterns, such as adhesions, this visually defined ground truth may therefore overestimate AI performance when compared with a histological reference standard. Computer vision methods based on machine learning can classically accomplish 4 types of tasks on an image: (i) classification (e.g., presence or absence of endometriosis); (ii) object detection (e.g., spatial localization of endometriosis by a bounding box); (iii) semantic segmentation (e.g., precise delimitation of an endometriosis lesion pixelwise); and (iv) instance segmentation (e.g., precise delimitation of different endometriosis lesions). This type of project requires the implementation and training of an artificial neural network. This training is necessarily guided by human identification of the structures to be recognized. In concrete terms, this will require surgeons to precisely identify and annotate endometriosis lesions on a massive number of laparoscopic images (typically, between 10 000 and 100 000). A simple, consensual, and exhaustive ontology is an essential prerequisite for this work. Notably, this massive number of images makes it unreasonable to attempt to histologically verify the presence of endometriosis for each lesion. Therefore, the veracity of the annotation will depend exclusively on the PPV of visual lesion identification. According to the present literature search, the PPV varies according to the type of lesion: excellent for endometriomas and black lesions but much weaker for adhesions and subtle lesions, which are not very specific for endometriosis. 21 , 23 , 26 , 29 , 62 These PPV values must be interpreted with caution. Histological confirmation is known to be imperfect: false negatives may result from superficial sampling, tissue degradation, or difficulties in histological interpretation. Even lesions considered highly specific, such as endometriomas or black peritoneal implants, may lack identifiable endometrial tissue. In this review, no predefined standards were applied regarding histological confirmation, as the included studies varied widely in their methods and reporting. This heterogeneity represents a limitation of our approach, which relied on the authors' statements of histological confirmation. Identifying deep endometriosis on the surface of organs is sometimes difficult, but it is nonetheless possible by indirect means involving three methods: retraction and/or deformation of the organs involved and obliteration of the spaces. This work combines a structured literature review with the construction of a visual ontology, offering a rigorous and reproducible framework for the annotation of endometriosis lesions. By integrating historical classifications with detailed visual descriptors, it lays a solid foundation for the development of AI‐based recognition tools. However, this approach has several limitations. The video database was used informally to verify the robustness of the proposed ontology and to generate illustrative images. However, this approach did not follow a formal methodological framework. A more rigorous validation, involving independent expert review and a data saturation process, could be considered in future work. Additionally, the majority of included studies were published between the late 1980s and early 2000s, reflecting historical perspectives that may not fully align with current surgical understanding. Moreover, the reliance on histological confirmation introduces a potential risk of false negatives, particularly in the case of subtle or superficial lesions. Finally, several included studies originated from the same authors or institutions, particularly among older publications, raising the possibility of overlapping patient populations. Although this could affect the reliability of aggregated quantitative data, we only reported ranges (minimum–maximum), thereby minimizing the impact of such potential redundancies. This search thus reveals predictable difficulties for visual identification without anatomopathological verification: (i) a predictable lack of veracity for certain nonspecific lesions, (ii) a reproducibility between annotators negatively impacted by a high number of classes and a differentiation between fragile classes, and (iii) a difficult delimitation of lesions whose contours may be poorly marked (subtle lesions, deep lesions), especially in the absence of tactile feedback such as palpation of indurations during laparoscopy. In addition to highlighting the foreseeable difficulties, this literature search confirms the relevance of the project. Three objectives emerge from the design of an endometriosis recognition algorithm: (i) to improve the PPV for the recognition of subtle lesions in particular; (ii) to achieve a diagnostic performance equivalent to that of an expert, regardless of the surgeon's level of experience; and (iii) to increase the completeness of lesion excision during the surgical procedure.

The present study established a systematic and relatively simple but complete ontology of the different visual aspects that endometriosis exhibits during laparoscopy. By synthesizing heterogeneous descriptions from the literature into a coherent framework, this ontology addresses a critical prerequisite for reproducible visual interpretation, expert annotation, and both clinical and research applications.

Laparoscopy allows for direct visual and histological assessment of endometriosis and is the only procedure that can definitively exclude the disease when imaging examinations are normal or inconclusive. 1 , 2 , 3 However, numerous studies have shown that the diagnostic performance and reproducibility of visual recognition of endometriosis during laparoscopy are poor. 4 , 5 , 6 , 7 , 8 The great diversity and subtle nature of the lesions, their small size, the variety of possible lesion locations, and the fact that they can be buried within organs or behind the peritoneum are all factors that limit their recognition by surgeons and the standardization of surgical procedures. 9 , 10 , 11 In addition, diagnostic accuracy is highly dependent on the surgeon's level of experience and training in endometriosis surgery, leading to substantial inter‐operator variability and further limiting reproducibility. 7 , 9 In this context, the use of artificial intelligence (AI), particularly machine learning, seems relevant for exhaustive lesion recognition and standardization of surgical procedures. The accuracy of image recognition using AI has dramatically increased in recent years. 12 In a recently published proof‐of‐concept study, our group demonstrated the feasibility of automatic visual recognition of endometriosis during laparoscopy using AI. 13 The implementation of such an AI project faces many obstacles, one of which is the absence of a reference system for the visual recognition of endometriosis lesions during laparoscopy. Standardization of the visual classification of lesions, in the form of an ontology (a class‐based organization with representation, formal naming, and definitions), is therefore an essential prerequisite to enable medical experts to annotate surgical data consistently and subsequently allow engineers to train and build an AI tool for endometriosis recognition. The objective of this study was to conduct a systematic search of the literature regarding the laparoscopic visual descriptions of endometriosis lesions and to propose a standardized visual ontology that can serve as a reproducible framework for expert annotation and for AI‐based lesion recognition.

Nicolas Bourdel is the CEO of SurgAR, and the other authors declare that they have no competing interests related to this study.

We conducted a systematic search of the literature using the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) 2020 criteria. 14 Two of the authors (A.N. and F.D.) independently searched the MEDLINE (via PubMed), EMBASE, and the Cochrane Library databases until May 2022, aiming to identify studies describing the laparoscopic visual appearance of superficial endometriosis, endometriomas, and deep infiltrating endometriosis. The complete search strategy—including all Boolean queries, the database‐specific adaptations, and the rationale for term selection—is given in Supplementary  Material 1 . To ensure comprehensiveness, we also manually screened the reference lists of all included articles for additional studies meeting the inclusion criteria. The protocol for this literature search was registered in the PROSPERO international prospective register of systematic reviews (Registration number: PROSPERO 2022 CRD42022354949). We selected articles published in English in scientific journals with guaranteed peer review and book chapters. Articles were only retained when their abstract clearly suggested that the authors aimed to provide a structured or intentional description of the laparoscopic appearance of endometriotic lesions. We included randomized controlled trials, prospective or retrospective cohort studies, literature reviews, and meta‐analyses. Case reports, case series, and letters to the editor were also included, provided that they reported relevant information. Two of the authors (A.N. and F.D.) independently conducted the first selection by eliminating articles whose titles or abstracts suggested that they dealt with themes too far from the scope. After this initial screening, the full texts of the remaining articles were examined, and all relevant articles and their references were carefully analyzed to identify any material possibly suitable for inclusion in the study. When available, positive predictive values (PPVs) were extracted or derived from the included studies as the proportion of visually identified lesions that were histologically confirmed as endometriosis. For each lesion type, PPVs were calculated based on the data reported by the original authors, using histological confirmation as the reference standard. Due to heterogeneity in study designs and reporting, PPVs are presented as ranges (minimum–maximum) rather than pooled estimates. The accumulated data in the literature concerning the visual appearance of the different forms of endometriosis was used to create an ontology that could be used for AI applications. The ontology is used to limit the complexity and organize the data into classes with representation, formal naming, and definitions. It is a formal, explicit description of classes and their properties in a domain of discourse. An ontology together with a set of individual instances of classes constitutes a knowledge base. 15 The aim of this way of organizing the data is to precisely define the parameters that allow the different types of endometriosis to be differentiated visually. This step is crucial in the context of machine learning for obtaining reasoned learning based on real knowledge. 16 To support the development of the present classification system, we relied on an international multicenter database of laparoscopic videos specifically created for our broader AI project (Institutional Review Board: 58723‐4/2016/EKU). This database includes prospectively recorded surgical procedures from four expert centers, with standardized exploratory sequences of the pelvic cavity. In the context of the present study, the video dataset was used for two main purposes: (i) to informally verify that each lesion described in the literature could be assigned to a specific class within the proposed ontology, ideally in a non‐ambiguous manner; and (ii) to extract representative still frames used to illustrate the different lesion types. A detailed description of the video database and its structure is provided in the companion article submitted alongside this manuscript.

Supplementary Material 1: Literature Search Strategy.