A method for analyzing bone texture from digital images.

Abstract

The invention relates to a method for analyzing bone texture, comprising: receiving an input X-ray image (6) showing an input bone, and a bone score analysis (11, 12) of the received input X-ray image by a bone score artificial intelligence (ANN1, ANN2) implemented by technical means. As a result of this bone score analysis, the bone score artificial intelligence gives: a global score depending at least on a trabecular bone score (TBS) depending on the texture of the trabecular bone part of the input bone shown in the received input X-ray image, and / or a trabecular bone score (TBS) depending on the texture of the trabecular bone part of the input bone shown in the received input X-ray image. The invention also relates to a corresponding device.

Description

[Technical field]

[0001] The present invention relates to a method for analyzing bone texture from a digital image.

[0002] The present invention also relates to an apparatus for analysing bone texture from a digital image.

[0003] The technical field of the present invention is typically that of deep learning, and in particular, but not limited to, methods for identifying individuals likely to be diagnosed with osteoporosis from appropriate screening of digital x-ray based images, who have impaired bone mass and bone microarchitecture as assessed by DXA BMD and TBS, or other equivalent methods. [Background technology]

[0004] In the early 1990s, the World Health Organization (WHO) conceptualized osteoporosis as a systemic skeletal disease characterized by reduced bone mass (quantity loss) and deterioration of bone tissue microarchitecture (quality loss), resulting in increased bone fragility and fracture risk (Non-Patent Document 1). Furthermore, in the early 2000s, the National Institutes of Health (NIH) defined osteoporosis as a skeletal disease with reduced bone strength and increased fracture risk (Non-Patent Document 2). Essentially, in osteoporosis, reduced bone strength leads to a traumatic outcome in the form of fragility fractures.

[0005] Bone strength reflects the integration of two major features: bone mass (i.e., bone density) and bone quality. Bone density is expressed as grams of mineral per unit area or volume, and in an individual is determined by the maximum bone mass and the amount of bone loss. Bone quality refers to bone structure, bone elasticity, turnover, accumulation of damage (e.g., microfractures), and mineralization. Bone structure is a general term used for many different entities and can be further subdivided. At the macro level, bone structure, known as bone macroarchitecture (also called bone macroarchitecture), represents the overall shape and geometric form of the bone, as well as its differentiation into cancellous bone (also called trabecular bone) and cortical bone (Non-Patent Document 3).

[0006] In any case, osteoporosis is characterized by fragility fractures. A fragility fracture is defined as a fracture that occurs due to a fall from a standing height in response to a mechanical force that would not normally cause a fracture (Non-Patent Document 4, Non-Patent Document 5). The proximal femur, spine, humerus, and forearm are the skeletal sites where fragility fractures occur most frequently. These fractures are called primary osteoporotic fractures.

[0007] Once fracture risk is identified, preventive measures are taken at a sort of "hierarchical" level: general lifestyle advice, a healthy lifestyle plus appropriate calcium and / or vitamin D supplementation, and / or pharmacotherapy. Based on the mechanism of action, there are two types of osteoporosis pharmacotherapy: antiresorptive agents (e.g., bisphosphonates, estrogen agonists / antagonists, estrogen, calcitonin, denosumab) that inhibit bone resorption, and osteogenic agents (e.g., teriparatide) that stimulate bone formation. Recently, romosozumab has been approved for its osteogenic effect (Non-Patent Document 6).

[0008] Over 9 million fragility fractures occur annually worldwide, and the number is expected to increase with the aging of the population (Non-Patent Document 7, Non-Patent Document 8). After the age of 50, it is estimated that one in two women and one in four men will suffer a major osteoporotic fracture in the remainder of their lives (Non-Patent Document 9). In women over 45 years of age, osteoporosis results in more days of hospitalization than many other diseases, including diabetes, myocardial infarction, and breast cancer. Furthermore, a previous fracture increases the risk of a subsequent fracture by 86%. Fractures are associated with high morbidity and mortality, and in older adults they are often a harbinger of disability, loss of independence, and premature death (Non-Patent Document 10). For example, overall, the cost of osteoporosis, including drug interventions, in the EU alone in 2010 was estimated at 37 billion euros, of which 66% accounted for the cost of treating accidental fractures, 5% for drug prevention, and 29% for long-term fracture care (Non-Patent Document 11). In addition, a report jointly prepared by the International Osteoporosis Foundation and the European Federation of Pharmaceutical Industries and Associations was published (Non-Patent Document 12).

[0009] Unfortunately, it is estimated that about 70% of people considered to be at risk for osteoporosis are never identified and are never even referred to an osteologist for proper diagnosis by DXA (the gold standard). Furthermore, many studies have shown that the percentage of patients who are registered and treated for osteoporosis, even after experiencing a hip fracture, has decreased significantly year by year. In fact, the majority (perhaps 80%) of high-risk individuals who have already experienced at least one osteoporotic fracture are not identified and are not treated (Non-Patent Document 13). In addition, a report jointly prepared by the International Osteoporosis Foundation and the European Federation of Pharmaceutical Industries and Associations (EFPIA) was reported (Non-Patent Documents 14 to 17).

[0010] There are many reasons for the underidentification of individuals at risk for osteoporosis and the decline in diagnosis and treatment of osteoporosis. Perhaps the biggest problem is the lack of awareness of bone disease among both the general public and health care professionals, many of whom do not understand the magnitude of the problem, much less how to prevent and treat bone disease. Even when patients experience fragility fractures, the significance of all osteoporotic fractures, including asymptomatic vertebral compression fractures, is not fully recognized and patients admitted with osteoporotic fractures are not reliably integrated into osteoporosis management plans to prevent a second fracture. Indeed, there is an international movement to develop multidisciplinary fracture liaison services (FLS) to address this latter issue and avoid subsequent second fragility fractures. FLS relies on the development of mechanisms and pathways to identify patients admitted to hospitals, emergency rooms, and urgent care clinics with osteoporotic fractures and to direct those patients to a well-designed osteoporosis management and treatment plan (Non-Patent Document 18).

[0011] At various ages before a fragility fracture occurs, there are many "red flags" that suggest potential problems with an individual's bone health. Although bone mineral density (BMD) testing still serves as the "gold standard" diagnostic test for identifying osteoporosis and fracture risk, population-wide BMD testing is not a cost-effective or practical method for assessing risk of bone disease. Although BMD testing is recommended for some populations (e.g., women over 65 years of age), for other individuals, i.e., the majority of people who do not have or are not at risk for bone disease, BMD testing is not used routinely. Widespread BMD testing makes little economic or medical sense. Rather, evidence supports evaluation of other risk factors first to identify subsets of high-risk individuals who are most likely to benefit from testing (e.g., young women with multiple risk factors, men and women with a history of fragility fractures, men and women with diseases that may substantially increase fracture risk, etc.). Some of these risk factors directly or indirectly affect BMD levels, while others (e.g., risk factors for falls) are independent of BMD. As such, many screening strategies based on risk assessment models have been developed. Unfortunately, several problems and limitations have slowed the development and widespread application of risk factor assessment tools. One of the key issues is related to the limitations of current medical knowledge about risk factors for bone health, but more importantly, the still-needed proactive screening behaviors related to awareness of both the general public and health care professionals (Non-Patent Documents 19-22). [Prior art documents] [Non-patent literature]

[0012] [Patent Document 1] FR2848694 [Patent Document 2] EP1576526A1 [Non-Patent Document 1] Consensus development conference:diagnosis,prophylaxis,and treatment of osteoporosis.Am J Med-1993-94,646-650 [Non-Patent Document 2] Osteoporosis prevention, diagnosis, and therapy.Jama.2001;285(6):785-95. [Non-Patent Document 3] Macro-and Microimaging of Bone Architecture,Engelke K et al.,in Principles of Bone Biology(Third Edition)Volume II,2008,Pages 1905-1942.Editors:John Bilezikian Lawrence Raisz T.John Martin-Publisher:Academic Press.Published Date:29th September 2008 [Non-Patent Document 4] Cummings SR et al.Epidemiology and outcomes of osteoporotic fractures.Lancet(London,England).2002;359(9319):1761-7. [Non-Patent Document 5] Warriner AH et al.,Minor,major,low-trauma,and high-trauma fractures:what are the subsequent fracture risks and how do they vary? Current osteoporosis reports.2011;9(3):122-8. [Non-Patent Document 6] Tu KN.,et al. Osteoporosis: A Review of Treatment Options. P&T: a peer-reviewed journal for formulary management. 2018;43(2):92-104.

Non-Patent Document 7

Non-Patent Document 8

Non-Patent Document 9

Non-Patent Document 10

Non-Patent Document 11

Non-Patent Document 12

Non-Patent Document 13

Non-Patent Document 14

Non-Patent Document 15

[0013] To overcome this need for aggressiveness and to avoid the cost overruns of systematic screening of the entire population, for example based on DXA, one can envision an appropriate approach that is automatic when appropriate. This latter approach could be performed, for example, on X-ray-based images performed annually on millions of individuals for reasons other than osteoporosis. Such an approach must be optimized in terms of false positives and should aim to increase the number of individuals at true risk of osteoporosis who are referred to a DXA facility for the "gold standard" diagnostic confirmation (assessment of both bone mass and bone quality).

[0014] It is an object of the present invention to provide a method or apparatus for rapid and simple analysis of bone texture and / or health that can be applied to images where bone mineral density (BMD) or trabecular bone score (TBS) cannot normally be obtained, i.e. images other than typically dual energy X-ray absorptiometry (DXA) images.

[0015] One aspect of the invention relates to a method (preferably computer implemented) for analysing the texture of a bone (preferably obtained by imaging from a digital image and selected in areas containing bone structure), comprising the following steps: - receiving an input digitized x-ray image indicative of an input bone; - a step of bone score analysis of the received input X-ray image by a bone score artificial intelligence implemented by technical means, the bone score artificial intelligence giving as a result of this bone score analysis: -- a global score that depends at least on a trabecular bone score (TBS) that depends on the texture (or on an empirical variogram of gray levels that quantifies the local variations in gray levels) of the trabecular bone portion of the input bone shown in the received input X-ray image, and / or -- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the input bone shown in the received input X-ray image, and / or a density score that depends on the bone density of the input bone shown in the received input X-ray image.

[0016] The bone score artificial intelligence can be a neural network.

[0017] The method according to the invention may comprise the following steps: - constructing a first training set by performing the following steps multiple times: --obtaining a first type of training image showing the cancellous bone portion of the training bone; --obtaining a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining by technical means from training images of a first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining by technical means a global score that depends on these density scores and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical variogram of gray levels) of the trabecular bone portion of the training bones shown in the training images of the first type; - training the Bone Score Artificial Intelligence by providing the Bone Score Artificial Intelligence with training images of the second type together with their associated ground truth including global scores determined for training images of the first type associated with the training images of the second type.

[0018] The method according to the invention may comprise the following steps: - constructing a second training set by performing the following steps multiple times: --obtaining a first type of training image showing the cancellous bone portion of the training bone; --obtaining a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining by technical means from training images of a first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type, and / or --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - training the Bone Score Artificial Intelligence by providing the Bone Score Artificial Intelligence with training images of the second type along with associated ground truth thereof, the training images including: -- density scores determined for the training images of the first type associated with this training image of the second type; and / or --a trabecular bone score determined for the training image of the first type associated with this training image of the second type.

[0019] The method according to the invention may comprise the following steps: - receiving an input X-ray image indicative of an input bone; - a first analysis step of the received input X-ray image by a first artificial intelligence implemented by technical means, the first artificial intelligence giving as a result of the first analysis a global score that depends on both: -- a density score depending on the bone density of the input bone shown in the received input x-ray image -- a Trabecular Bone Score (TBS) that depends on the texture (or quantifies the local variation in grey levels from an empirical grey level variogram) of the trabecular bone portion of the input bone shown in the received input X-ray image; - a second analysis step of the received input X-ray image by a second artificial intelligence implemented by technical means, the second artificial intelligence providing as a result of the second analysis: -- a density score dependent on the bone density of the input bone shown in the received input X-ray image, and / or -- a Trabecular Bone Score (TBS) that depends on the texture (or quantifies the local variation in grey levels from an empirical grey level variogram) of the trabecular bone portion of the input bone shown in the received input X-ray image; - a third analysis step by a third artificial intelligence implemented by technical means, the third artificial intelligence having as input the results of the first and second analyses and having as output a result that depends on the consistency between the result of the first analysis and the result of the second analysis.

[0020] The third artificial intelligence may use as further input at least one of the following parameters: the age of the patient from whom the received input X-ray image was acquired, the gender of the patient from whom the received input X-ray image was acquired, the body type of the patient from whom the received input X-ray image was acquired, the type of equipment from which the received input X-ray image was acquired, and / or acquisition parameters of the equipment from which the received input X-ray image was acquired.

[0021] The first artificial intelligence and / or the second artificial intelligence may be a neural network.

[0022] The first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be three different artificial intelligences.

[0023] The technical means for implementing the first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be the same technical means.

[0024] The method according to the invention may comprise the following steps: - constructing a first training set by performing the following steps multiple times: --obtaining a first type of training image showing the cancellous bone portion of the training bone; --obtaining a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining by technical means from training images of a first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining by technical means a global score that depends on these density scores and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation of gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the training images of the first type; - training the first artificial intelligence by providing the first artificial intelligence with training images of the second type together with their associated ground truth including global scores determined for training images of the first type associated with the training images of the second type.

[0025] The method according to the invention may comprise the following steps: - constructing a second training set by performing the following steps multiple times: --obtaining a first type of training image showing the cancellous bone portion of the training bone; --obtaining a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining by technical means from training images of a first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type, and / or --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - training a second artificial intelligence by providing the second artificial intelligence with training images of the second type together with associated ground truth including: -- density scores determined for the training images of the first type associated with this training image of the second type; and / or --a trabecular bone score determined for the training image of the first type associated with this training image of the second type.

[0026] The first artificial intelligence and the second artificial intelligence may be trained using the same database of training images of the first type and training images of the second type.

[0027] The method according to the invention may comprise the following steps: constructing a third training set by performing the following steps multiple times: --obtaining a first type of training image showing the cancellous bone portion of the training bone; -- obtaining a second type of training image, which is an X-ray based image, showing the same training bone; -- determining by technical means from training images of a first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining by technical means a global score which depends on this density score and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical variogram of gray levels) of the trabecular bone portion of the training bones shown in the training images of the first type; -- carrying out first and second analyses by first and second artificial intelligences on the second type of training images; and - training a third artificial intelligence by learning from the difference between the scores obtained from the training images of the first type and the scores obtained from training images of the second type of the same training bones.

[0028] The first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be trained separately.

[0029] The first type of training image and the second type of training image may be acquired of the same training bone and are acquired less than six months apart.

[0030] The first type of training image may be a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0031] The second type of training image is preferably not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0032] The received input X-ray image is preferably not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0033] The received input x-ray image may be a digital x-ray image having a spatial resolution of less than 1 mm per pixel.

[0034] Another aspect of the invention relates to a computer program comprising instructions for carrying out the steps of the method according to the invention when executed by a computer.

[0035] Another aspect of the invention relates to a computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to the invention.

[0036] Another aspect of the invention relates to a computer-readable storage medium containing instructions which, when executed by a computer, cause the computer to perform the steps of the method according to the invention.

[0037] Another aspect of the invention relates to a device for analysing bone texture (preferably obtained by imaging from a digital image and selected in areas containing bone structures), comprising: - means arranged and / or programmed and / or configured to receive an input digitized X-ray image indicative of an input bone; - technical means arranged and / or programmed and / or configured to implement a bone score artificial intelligence arranged and / or programmed and / or configured to implement a bone score analysis of a received input X-ray image, wherein the bone score artificial intelligence is arranged and / or programmed and / or configured to provide as a result of this bone score analysis: -- a global score that depends at least on a trabecular bone score (TBS) that depends on the texture (or on an empirical variogram of gray levels that quantifies the local variations in gray levels) of the trabecular bone portion of the input bone shown in the received input X-ray image, and / or -- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the input bone shown in the received input X-ray image, and / or a density score that depends on the bone density of the input bone shown in the received input X-ray image.

[0038] The bone score artificial intelligence can be a neural network.

[0039] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to construct a first training set by performing the following steps a plurality of times: - acquiring (by technical means for acquiring) a first type of training image showing the cancellous bone portion of the training bone; -- acquiring (by technical means for acquiring) a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining, by technical means for determining, from the training images of the first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining, by technical means for determining, a global score which depends on these density scores and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - means arranged and / or programmed and / or configured to train the Bone Score Artificial Intelligence by providing to the Bone Score Artificial Intelligence training images of a second type together with their associated ground truth including global scores determined for training images of the first type associated with the training images of the second type.

[0040] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to construct a second training set by performing the following steps a plurality of times: - acquiring (by technical means for acquiring) a first type of training image showing the cancellous bone portion of the training bone; -- acquiring (by technical means for acquiring) a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining, by technical means for determining, from the training images of the first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type, and / or --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - means arranged and / or programmed and / or configured to train a bone-scoring artificial intelligence by providing the artificial intelligence with training images of a second type together with its associated ground truth including: -- density scores determined for the training images of the first type associated with this training image of the second type; and / or --a trabecular bone score determined for the training image of the first type associated with this training image of the second type.

[0041] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to receive an input X-ray image indicative of an input bone; - technical means arranged and / or programmed and / or configured to implement a first artificial intelligence arranged and / or programmed and / or configured to implement a first analysis of a received input X-ray image, the first artificial intelligence arranged and / or programmed and / or configured to give as result of the first analysis a global score which is dependent on both: -- a density score depending on the bone density of the input bone shown in the received input x-ray image -- a Trabecular Bone Score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the input bone as shown in the received input X-ray image; - technical means arranged and / or programmed and / or configured to implement a second artificial intelligence arranged and / or programmed and / or configured to implement a second analysis of the received input X-ray image, wherein the second artificial intelligence is arranged and / or programmed and / or configured to provide as a result of the second analysis: -- a density score dependent on the bone density of the input bone shown in the received input X-ray image, and / or -- a Trabecular Bone Score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the input bone as shown in the received input X-ray image; - technological means arranged and / or programmed and / or configured to implement a third artificial intelligence arranged and / or programmed and / or configured to implement a third analysis, wherein the third artificial intelligence has as input the results of the first and second analyses and has as output a result that depends on the consistency between the result of the first analysis and the result of the second analysis.

[0042] The third artificial intelligence may be arranged and / or programmed and / or configured to use as further input at least one of the following parameters: the age of the patient from whom the received input X-ray image was obtained, the gender of the patient from whom the received input X-ray image was obtained, the body type of the patient from whom the received input X-ray image was obtained, the type of equipment from which the received input X-ray image was obtained, and / or acquisition parameters of the equipment from which the received input X-ray image was obtained.

[0043] The first artificial intelligence and / or the second artificial intelligence may be a neural network.

[0044] The first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be three different artificial intelligences.

[0045] The technical means for implementing the first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be the same technical means.

[0046] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to construct a first training set by performing the following steps a plurality of times: - acquiring (by technical means for acquiring) a first type of training image showing the cancellous bone portion of the training bone; -- acquiring (by technical means for acquiring) a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining, by technical means for determining, from the training images of the first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining, by technical means for determining, a global score which depends on these density scores and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - means arranged and / or programmed and / or configured to train a first artificial intelligence by providing said first artificial intelligence with training images of a second type together with their associated ground truth comprising global scores determined for training images of the first type associated with said training images of the second type.

[0047] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to construct a second training set by performing the following steps a plurality of times: - acquiring (by technical means for acquiring) a first type of training image showing the cancellous bone portion of the training bone; -- acquiring (by technical means for acquiring) a related second type of training image, which is an X-ray based image, showing the same training bone. -- determining, by technical means for determining, from the training images of the first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type, and / or --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. - means arranged and / or programmed and / or configured to train a second artificial intelligence by providing said second artificial intelligence with training images of a second type together with its associated ground truth comprising: -- density scores determined for the training images of the first type associated with this training image of the second type; and / or --a trabecular bone score determined for the training image of the first type associated with this training image of the second type.

[0048] The first artificial intelligence and the second artificial intelligence may be arranged and / or programmed and / or configured to train using the same database of training images of the first type and training images of the second type.

[0049] An apparatus according to the invention may include: - means arranged and / or programmed and / or configured to construct a third learning set by performing the following steps a plurality of times: - acquiring (by technical means for acquiring) a first type of training image showing the cancellous bone portion of the training bone; -- acquiring (by technical means for acquiring) a second type of training image, which is an X-ray based image, showing the same training bone. -- determining, by technical means for determining, from the training images of the first type: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- A trabecular bone score (TBS) that depends on the texture (or quantifies the local variation in gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the first type of training image. determining, by technical means for determining, a global score which depends on this density score and the cancellous bone score from: --- a density score that depends on the bone density of the training bones shown in the training images of the first type; and --- a trabecular bone score (TBS) that depends on the texture (or quantifies the local variation of gray levels from an empirical gray level variogram) of the trabecular bone portion of the training bones shown in the training images of the first type; - performing first and second analyses by the first and second artificial intelligence on the second type of training images; and - means arranged and / or programmed and / or configured to train a third artificial intelligence by learning from the difference between scores obtained from training images of the first type and scores obtained from training images of a second type of the same training bone.

[0050] The first artificial intelligence, the second artificial intelligence, and the third artificial intelligence may be arranged and / or programmed and / or configured to train separately.

[0051] The means arranged and / or programmed and / or configured to train the first artificial intelligence and the means arranged and / or programmed and / or configured to train the second artificial intelligence may be arranged and / or programmed and / or configured together to check that the first type of training image and the second type of training image are acquired of the same training bone and are acquired less than six months apart.

[0052] The first type of training image may be a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0053] The second type of training image is preferably not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0054] The received input X-ray image is preferably not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0055] The received input x-ray image may be a digital x-ray image having a spatial resolution of less than 1 mm per pixel.

[0056] Other advantages and features of the present invention will become apparent upon consideration of the detailed description of non-limiting embodiments and the accompanying drawings. [Brief description of the drawings]

[0057] [Figure 1] FIG. 1 shows some steps of a first embodiment (best realized mode) of a method 100 according to the invention at the industrialization stage, i.e. a module of an artificial neural network (ANN) model in a final product of a first embodiment (best realized mode) of a device according to the invention implementing the first embodiment of a method 100 according to the invention. [Diagram 2] FIG. 2 illustrates the principle of the global score 16 in this embodiment of the method 100 according to the invention in a table of classification of bone health (and associated risk) using bone density x bone texture (e.g. minimum BMD T-score x TBS) and further below shows bone health categories based on fracture risk. [Diagram 3] FIG. 3 shows some steps of a first embodiment of a method 100 according to the invention, during the repository extraction phase and during the ground truth processing phase. [Figure 4] FIG. 4 illustrates some steps of a first embodiment of a method 100 according to the invention, during the repository extraction phase. [Diagram 5] FIG. 5 shows some steps of a first embodiment of a method 100 according to the invention, in the pre-processing stage and in the deep learning or training stage. [Figure 6] FIG. 6 shows some steps of a first embodiment of a method 100 according to the invention, in a testing and clinical optimization phase. [Figure 7] FIG. 7 shows some steps of a first embodiment of a method 100 according to the invention, in a validation and industrialization phase. [Figure 8] FIG. 8 shows an example of a variogram V on a log-log scale with parameters a, b, c, d, e used in the learning step of the method 100. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0058] These embodiments are in no way limiting, and variations of the invention can be considered that include only a selection of the features or steps described or illustrated below, and that this selection of features or steps may be separated from other described or illustrated features or steps (even if this selection is taken from a sentence that includes these other features or steps), if that portion is sufficient to provide a technical advantage or to distinguish the invention over the prior art. This selection may include at least one feature, preferably a functional feature, without structural details or with only a portion of structural details, if that portion is sufficient to provide a technical advantage or to distinguish the invention over the prior art.

[0059] 1 to 8, a first embodiment of a computer-implemented method 100 according to the present invention will now be described.

[0060] 1 to 7 show different parts of this method 100.

[0061] These figures are linked to each other.

[0062] for example, Link 1 in Figure 3 corresponds to Link 1 in Figure 5. Link 2 in Figure 3 corresponds to Link 2 in Figure 5. Link 3 in Figure 4 corresponds to link 3 in Figure 5. Link 4 in FIG. 4 corresponds to link 4 in FIG. Link 5 in FIG. 4 corresponds to link 5 in FIG. Link 7 in FIG. 4 corresponds to link 7 in FIG. link 8 in FIG. 4 corresponds to link 8 in FIG. link 10 in FIG. 3 corresponds to link 10 in FIGS. Link 33 in FIG. 5 corresponds to link 33 in FIG. Link 44 in FIG. 5 corresponds to link 44 in FIG.

[0063] Method 100 is intended to assist radiologists in identifying individuals at potential high risk for osteoporosis, defined as having reduced bone mineral density (BMD) and bone microarchitecture, who may be referred to specialized bone disease centers or DXA centers for confirmation of the diagnosis and / or appropriate disease management.

[0064] There are two types of bone tissue in the human skeleton: cortical bone and cancellous bone. The two types of bone are macroscopically and microscopically different, but have identical chemical composition. Trabecular bone, also called cancellous or spongy bone, is a highly porous bone (typically 75-95% porosity) that encloses numerous large spaces, forming a honeycomb-like or spongy network of cancellous bone with large areas of remodeling and a high turnover rate. More precisely, the bone matrix is ​​organized into a three-dimensional lattice-like structure of interconnected rods and plates called trabeculae, which surround pores filled with bone marrow, fat, and blood vessels. This cancellous bone is usually aligned along the direction of mechanical stress. Trabecular bone is usually surrounded by a shell of compact bone (also called cortical bone), which provides greater strength and stiffness. Trabecular bone accounts for approximately 20% of the bone mass in the human skeleton, but accounts for approximately 80% of the bone remodeling activity, providing structural support and flexibility, making this bone the most suitable target for therapies affecting bone metabolism, as well as the first bone to be affected by metabolic disorders such as primary and secondary osteoporosis.

[0065] Thus, cancellous bone (also called trabecular bone) is a part of bone well known to those skilled in the art, and is abundant in short bones such as vertebrae. It is the main component of the vertebral bodies, carpal bones and the central part of long bones. It is weak and consists of non-concentrically arranged platelets or trabeculae, surrounding a cavity or cavity filled with red bone marrow. It is therefore - Composed of bone platelets or trabeculae giving it a spongy appearance; - It is less strong than cortical bone, but provides three-dimensional resistance to stresses; - Present mainly in the central part of the bone; - rich in blood vessels and undergoing constant remodeling to adapt the bone structure to the frequency and intensity of impacts; - Prone to compression fractures under intense pressure and susceptible to osteoporosis.

[0066] The method 100 suitably analyzes digital x-ray images acquired during routine clinical practice from a Picture Archiving and Communication System (PACS) or cloud-based system. These x-ray images are primarily acquired for reasons other than osteoporosis. Individuals scanned during this routine practice are not typically diagnosed with osteoporosis.

[0067] The method 100 is a global ensemble model that uses a combination of multiple artificial neural networks (ANNs), ANN1 and ANN2, to perform analysis of digital x-ray images.

[0068] These ANNs are trained using bone mineral density (BMD) assessed by dual-energy X-ray absorptiometry (DXA) and trabecular bone score (TBS, a textural parameter that correlates with bone microarchitecture; see US Pat. No. 5,399,146) as ground truth. The combination of BMD and TBS currently constitutes the gold standard for osteoporosis diagnosis.

[0069] The outputs of multiple ANNs are combined to provide a final risk assessment.

[0070] Using this method, method 100 performs appropriate screening, either as a systematic background task via a PACS system or as an active push via a cloud-based platform, to identify individuals at high risk for osteoporosis.

[0071] Therefore, the global approach of this method100 is to use bone-related x-ray images from the PACS to appropriately screen patients in a background task manner (or as an active push via a cloud-based platform) to identify individuals most likely to be at high or very low risk for osteoporosis, as defined by bone mineral density (BMD) by DXA and trabecular bone score (TBS).

[0072] For those identified as high risk, a comprehensive automated report can optionally be generated, suggesting referral to a bone specialist or DXA centre for diagnostic confirmation (BMD+TBS).

[0073] The approach of method 100 is based on a combination of supervised deep learning models. Artificial Intelligence (AI) models ANN1 and ANN2 of method 100 are trained on ground truth DXA (but not limited to) data (BMD+TBS) to process digital X-ray based images.

[0074] The method 100 is optimized for clinical outcomes (e.g., low false positive rates), short processing times, and seamlessly integrates into the radiology workflow.

[0075] The method 100 includes the following steps, in the following order: - The repository extraction stage (shown in Figure 3 (left) and Figure 4), - The ground truth processing stage (shown in Figure 3 (right)), - a pre-treatment stage (shown in Figure 5 (left)), - the deep learning or training phase (illustrated in Figure 5 (right)), - the test phase (shown in Figure 6 (left)), - Clinical optimization phase (shown in Figure 6 (right)), - the verification stage (shown in Figure 7 (left)), - The industrialization stage (shown in Figure 7 (right part)).

[0076] As shown in FIG. 1, the method 100 acquires and receives (at the final industrialized stage in FIG. 1 and on the right in FIG. 7) an input digitized X-ray image 6 showing an input bone.

[0077] The received input X-ray image 6 is preferably not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

[0078] In this embodiment, each image 6, 9 or 19 is a digitized image.

[0079] The received input x-ray image 6 is preferably a digital x-ray image having a spatial resolution of less than 1 mm per pixel.

[0080] The method 100 then comprises a first analysis 11 of the received input X-ray image 6 by a first artificial intelligence ANN1 implemented by technical means (typically comprising or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated), and / or one microprocessor (preferably dedicated), and / or software means). The first artificial intelligence gives, as a result of the first analysis, a global score 16, shown in FIG. 2, depending both on: - a density score that depends on (or consists of) the bone density of the input bone shown in the received input X-ray image - Trabecular Bone Score (TBS) which depends on the texture of the trabecular bone portion of the input bone as shown in the received input X-ray image

[0081] The global score 16 preferably depends only on this density score (preferably composed of bone density) and the trabecular bone score (TBS), and is therefore preferably a combination of this density score (preferably composed of bone density) and the trabecular bone score (TBS).

[0082] Nevertheless, the ANN 1 does not determine or calculate density scores or TBS from the images 6 but directly determines a global score 16 .

[0083] The first artificial intelligence is the neural network.

[0084] GlobalScore16 is a multi-class classifier.

[0085] The global score 16 has a finite number of possible values. It has only nine possible values ​​as shown in Figure 2. Each of these values ​​is a positive integer. Each of these values ​​is preferably a positive integer among 1, 2, 3, 4, 5, 5, 7, 8, and 9.

[0086] In all of the illustrations of Figures 1 to 8, the density scores can be as follows: - Equal to bone density - Determined according to the value of BMD. For example, -- If BMD T value ≤ -2.5, density score = "osteoporosis" or "osteoporotic" -- If BMD T value is ]-2.5,-1[, density score = "osteopenic" or "osteopenic" -- If BMD T value ≥ -1, density score = "normal" bone mineral density is determined from the images 9 according to conventional prior art techniques. Bone mineral density (BMD) is typically determined from the absorption of each beam by the bone. Dual-energy X-ray absorptiometry is the most widely used and most thoroughly studied bone densitometry technique.

[0087] Typically, bone density is obtained directly from dual-energy x-ray absorptiometry (DXA) images and typically may be metadata of the dual-energy x-ray absorptiometry (DXA) image itself.

[0088] The Trabecular Bone Score (TBS) is a texture parameter that quantifies the local variation in gray levels and is derived from the evaluation of an empirical variogram of gray levels in a digitized image, which is typically a dual-energy X-ray absorptiometry (DXA) image, but may also be other digital X-ray images or many other X-ray imaging modalities.

[0089] A more detailed description and examples of TBS can be found in Non-Patent Document 23.

[0090] Nevertheless, in this particular case, the global score 16 is determined by the first artificial intelligence without calculation of the TBS and without calculation or determination of an empirical variogram of the grey levels of the received input X-ray image 6.

[0091] The method 100 then comprises a second analysis 12 of the received input X-ray image 6 by a second artificial intelligence ANN2 implemented by technical means (typically comprising or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated), and / or one microprocessor (preferably dedicated), and / or software means). The second artificial intelligence gives as a result of the second analysis: - a density score that depends on (or consists of) the bone density of the input bone shown in the received input X-ray image 6, and / or - a trabecular bone score (TBS) that depends on the texture of the trabecular bone portion of the input bone shown in the received input X-ray image 6

[0092] The second type of artificial intelligence is neural networks.

[0093] The second type of artificial intelligence is the regression model, which gives or infers continuous values.

[0094] The TBS is determined by the second artificial intelligence without calculation or determination of an empirical variogram of the grey levels of the received input X-ray image 6.

[0095] The method 100 then comprises a third analysis 13 by a third artificial intelligence AI3 implemented by technical means (typically including or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated), and / or one microprocessor (preferably dedicated), and / or software means). The third artificial intelligence has as input the results of the first and second analysis and has as output a result that depends on the consistency of the result of the first analysis with the result of the second analysis. Typically, the output of AI3 depends on the consistency of the result of the first analysis of ANN1 with the result of the second analysis of ANN2 using weighting information from patient-related metadata 18 (soft tissue thickness, age, BMI, etc.). - Case #1: ANN1 = 6, 7, 8 or 9 (global score) ANN2 = BMD status osteopenic or osteoporotic, texture status partially degraded or degraded AI3 result = Flag - Case #2: ANN1 = 6, 7, 8 or 9 (global score) ANN2 = normal bone density status, partially degraded or degraded texture status AI3 result = weighted response to evaluate whether to flag or not depending on metadata 18 - Case #3: ANN1 = 6, 7, 8 or 9 (global score) ANN2 = BMD status osteopenic or osteoporotic, texture status normal AI3 result = weighted response to evaluate whether to flag or not depending on metadata 18 - Case #4: ANN1 = 1,2,3,4 or 5 (global score) ANN2 = Normal bone density status, normal texture status AI3 result = No flag - Case #5 ANN1 = 1,2,3,4 or 5 (global score) ANN2 = BMD status osteopenic or osteoporotic, texture status normal AI3 result = weighted response to evaluate whether to flag or not depending on metadata 18 - Case #6 ANN1 = 1,2,3,4 or 5 (global score) ANN2 = Bone density status, texture status partially degraded or degraded AI3 result = weighted response to evaluate whether to flag or not depending on metadata 18

[0096] The third artificial intelligence uses metadata 18 as a further input, the metadata 18 including at least one of the age of the patient from whom the received input X-ray image was obtained, the gender of the patient from whom the received input X-ray image was obtained, the body type of the patient from whom the received input X-ray image was obtained, the type of equipment from which the received input X-ray image was obtained, and / or the acquisition parameters of the equipment from which the received input X-ray image was obtained.

[0097] The third artificial intelligence may be a neural network, but is preferably a binary classifier such as a decision tree, a random forest, or a support vector machine.

[0098] Thus, as a final step, the outputs of the two aforementioned approaches ANN1 and ANN2, together with additional selected metadata18 (e.g., age, sex, body type, type of equipment, acquisition parameters), are input as input variables of AI3, which is typically a learning model like a decision tree. The classification tree provides as output the best combination for the final classification optimization (likelihood to be selected as osteoporosis with deteriorated bone microarchitecture, assessed by DXA BMD and TBS). Specific techniques, such as bagging decision trees (also called ensemble methods), are used to consider the likelihood of multiple image sets for a given time point and a given individual.

[0099] First AI, second AI, and third AI are three different AIs or three different AI architectures.

[0100] The technical means for implementing the first artificial intelligence, the second artificial intelligence and the third artificial intelligence are preferably integrated in the same technical means, i.e. but not limited to, preferably within the same at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated) and / or one microprocessor (preferably dedicated) and / or in an embedded module by software means.

[0101] These artificial intelligences ANN1, ANN2, and AI3 are trained in advance.

[0102] Thus, the method 100 includes, prior to the industrialization stage analysis steps 11, 12 and 13, the following steps: - constructing a first training set by performing the following steps multiple times: --obtaining (during the repository extraction phase) training images 9 of a first type (of a first training set) showing the cancellous bone portion of the training bones; -- obtaining (during the repository extraction phase) related second type 19 training images (of the first training set) that are X-ray based images showing the same training bone, but not necessarily showing its cancellous portion; -- determining (during the ground truth processing phase) from training images 9 of a first type of a first training set by technical means (typically comprising or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated) and / or one microprocessor (preferably dedicated) and / or by software means) (but without implementing artificial intelligence on the images 9): --- a density score 14 that depends on or is equal to the bone density of the training bones shown in the first type of training images of the first training set. Typically, the bone density is determined directly from the first type of training images as DXA metadata of the first type of training images; and --- A trabecular bone score (TBS) 15 that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type of the first training set. Typically, this TBS 15 is determined by calculating or determining an empirical variogram of the grey levels of the training images of the first type (Patent Document 2). -- determining (during the ground truth processing stage) a global score 16 that depends (preferably only depends) on these density scores and the trabecular bone scores (but without implementing any artificial intelligence on the images 9, and / or on the pre-determined density scores 14, and / or on the pre-determined TBS 15), by technical means (typically comprising or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated) and / or one microprocessor (preferably dedicated) and / or by software means), from: --- a density score 14 that depends on or is equal to the bone density of the training bones shown in the training images 9 of the first type of the first training set, and --- a trabecular bone score (TBS) 15 that depends on the texture of the trabecular bone portion of the training bone shown in the training image 9 of the first type of the first training set; for example, Depending on the value of bone mineral density (BMD), a density score or status can be determined. For example, -- If BMD T value ≤ -2.5, density score = "osteoporosis" or "osteoporotic" -- If BMD T value is ]-2.5,-1[, density score = "osteopenic" or "osteopenic" -- If BMD T value ≥ -1, density score = "normal" - The values ​​shown below are examples, since the setting algorithm is adapted to technical and clinical information, but the bone texture status can be determined depending on the TBS value. -- If TBS>1.31, then bone texture status = "normal" -- If TBS is [1.23,1.31], bone texture status = "partially degraded" -- If TBS<1.23, bone texture status = "degraded" - The global score 16 can be determined according to the following rules (as shown in Figure 2): -- If density score = "normal" and bone texture status = "normal", then global score = 1 -- If density score = "normal" and bone texture status = "partially degraded", then global score = 2 -- If density score = "normal" and bone texture status = "degraded", then global score = 3 -- If density score = "osteopenic" or "osteopenic" and bone texture status = "normal", then global score = 4 -- If density score = "osteopenic" or "osteopenic" and bone texture status = "partially degraded", then global score = 5 -- If density score = "osteopenic" or "osteopenic" and bone texture status = "deteriorated", then global score = 6 -- If density score = "osteoporotic" or "osteoporotic" and bone texture status = "normal", then global score = 7 -- If density score = "osteoporotic" or "osteoporotic" and bone texture status = "partially degraded", then global score = 8 -- If density score = "osteoporotic" or "osteoporotic" and bone texture status = "deteriorated", then global score = 9 - training (during the pre-processing and deep learning or training phases) a first artificial intelligence ANN 1 by providing it with training images 19 of the second type of the first training set (step 7) together with its associated ground truth 1 comprising or consisting of a global score 16 determined for training images 9 of the first type of the first training set. --ANN1 is (but is not limited to) a MLP, CNN with multiple layers (e.g., 50 layers or more). In a preferred embodiment, ANN1 is a multi-class classifier convolutional neural network. -- This learning step is done by gradient backpropagation by optimizing the accuracy (TP / (TP+FP)) (TP=true positives, FP=false positives). -- Training is continued until the minimum test loss (categorical cross-entropy loss function) is reached while avoiding overfitting.

[0103] For the first training set, the first type of training images 9 and the second type of training images 19 are acquired of the same training bone and are acquired less than six months apart.

[0104] For the first training set, the first type of training images 9 are dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

[0105] For the first training set, the second type of training images 19 are, for example, DICOM (short for "Digital Imaging and Communications in Medicine") images, rather than dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

[0106] For the first training set, the second type of training images 19 are preferably digital X-ray images having a spatial resolution of less than 1 mm per pixel.

[0107] The method 100 also includes, prior to the analysis steps 11, 12, and 13, the following steps. - constructing a second training set by performing the following steps multiple times: --obtaining (during the repository extraction phase) first type 9 training images (of the second training set) showing cancellous bone portions of training bones. --obtaining (during the repository extraction phase) related second type 19 training images (of the second training set) that are X-ray based images showing the same training bone but not necessarily showing its cancellous portion. -- determining (during the ground truth processing phase) from training images 9 of the first type of the second training set by technical means (typically comprising or consisting of at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated) and / or one microprocessor (preferably dedicated) and / or by software means) (but without implementing any artificial intelligence on the images 9): --- a density score 14 that depends on or is equal to the bone density of the training bones shown in the first type training images 9 of the second training set. Typically, the bone density is determined directly from the first type training images as metadata of the DXA of the first type training images; and / or --- A trabecular bone score (TBS) 15 that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type 9 of the second training set. Typically, this TBS 15 is determined by calculating or determining an empirical variogram of the grey levels of the training images of the first type (Patent Document 2). - training (during a pre-processing phase and a deep learning or training phase) a second artificial intelligence ANN2 by providing (step 8) to the second artificial intelligence training images 19 of a second training set of a second type together with its associated ground truth 2 comprising or consisting of: a density score 14 determined for a training image 9 of a first type of the second training set associated with a training image 19 of a second type of this second training set, and / or a cancellous bone score 15 determined for a training image 9 of a first type of the second training set associated with a training image 19 of a second type of this second training set.

[0108] In a preferred embodiment of the method 100, - ANN2 is (but is not limited to) a MLP, CNN with multiple layers (e.g., 50 layers or more). In a preferred embodiment, ANN1 is a recurrent convolutional neural network. - This learning step is done by gradient backpropagation by minimizing a quadratic loss function until the minimum test loss is reached while avoiding overfitting. - This training continues until the minimum test loss (quadratic loss function) is reached while avoiding overfitting.

[0109] For the second training set, the first type of training images and the second type of training images are acquired of the same training bone and are acquired less than six months apart.

[0110] For the second training set, the first type of training images are dual-energy x-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

[0111] For the second training set, the second type of training images are, for example, DICOM images, rather than dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

[0112] For the second training set, the second type of training images 19 are preferably digital X-ray images having a spatial resolution of less than 1 mm per pixel.

[0113] The determined cancellous bone score 15 is a texture parameter that quantifies the local variation of grey levels and is derived from the evaluation of an empirical variogram of grey levels of a first type of digitized image 9. This digitized image 9 is typically a dual energy X-ray absorptiometry (DXA) image.

[0114] For these steps for training the first artificial intelligence ANN1 and for these steps for training the second artificial intelligence ANN2, a cancellous bone score (TBS) 15 is determined from each training image 9 of the first type without implementing artificial intelligence on the images 9, typically according to the method described in Patent Document 2 or according to the following steps a) to i) of the training method.

[0115] a) Determine the optimized pixel sampling S of the image 9. For each pixel P i =(x i ,y i )∈S is its gray level value h(P i ). From image 9, (step d)) the variogram V i Determine the pixel sampling S to select the locations where is evaluated. These locations are optimized so that the final TBS is adapted and efficient for a given clinical context and anatomical site. In the most basic sense, the sampling region corresponds to the area in the image where bone is present. Depending on the bone site, a subsample of this region where bone is present in the image can be used to improve performance.

[0116] b) for at least one region of interest (ROI) of pixel sampling S, preferably for a plurality of ROIs, - Depending on the given region of interest (ROI), determine the calculation distance R0. The range R0 is determined depending on the skeletal site and the image resolution of the image 9. In DXA systems, the value of R0 is between 1 cm and 2 cm. - Selecting a set I of predefined directions according to a given region of interest (ROI). This step of selecting the set I of predefined directions is carried out by using N directional unit vectors

[0117]

number

[0118] (N is a positive integer and θ k ∈[-π,π],∀k∈1,…,N,θ k is the vector around the pixel under consideration.

[0119]

number

[0120] (The angle at which the Typically, N>2. The maximum value of N depends on the complexity of the bone structure of the bone being considered in the ROI and its image resolution. Typically, N<9 for bones with non-complex bone structure such as the vertebrae or lumbar spine, but for complex structures such as the proximal femur, N may be significantly increased. For example, N=3, 4, 6, or 8. The N direction vectors are preferably uniformly distributed with an angle 2π / N around the pixel under consideration. Given a set of directions I(vector

[0121]

number

[0122] ) depends on: - a bony skeletal site on the image and / or on the ROI (the human or animal tissue is bone), and / or - the region of interest (ROI) to be considered, and / or - Image resolution, and / or - Image signal-to-noise ratio. The determination of this direction set I is - Improve the ability to identify or predict fractures; and / or - allowing better measurement reproducibility (or precision), and / or - Allows better correlation with bone microstructure. Therefore, in the next step d), the variogram VPi To calculate the value h(P i ) to pixels located along a line in a particular direction. These directions depend on both the type of bone (i.e., skeletal site) and the region of interest selected for measurement on this bone to optimize the texture measurement. Indeed, the chosen directions are related to the bone morphology, especially the direction of the trabeculae in cancellous bone. For example, in the spine, the preferred direction of the trabeculae is vertical, so the chosen directions are vertical and horizontal [-Π / 2,0,Π / 2,Π] (parallel and perpendicular to the orientation of the trabeculae).

[0123] c) For each pixel Pi=(x i ,y i )∈S and each direction

[0124]

number

[0125] About

[0126]

number

[0127] Move along the line by a distance r ∈ [1,R0] (in pixels). Change the gray value of such a pixel to

[0128]

number

[0129] This is expressed as:

[0130] d) For each pixel, calculate at least one variogram of the gray levels of the sampling S as a function of the distance r along these directions I (i.e., by moving at least one distance r ∈ [1,R0] from this pixel. As mentioned before, moving a distance r is the sum of the distances of each pixel P i =(x i ,y i )∈S and each direction

[0131]

number

[0132] About

[0133]

number

[0134] This is done by moving a distance r ∈ [1, R0] pixels along

[0135]

number

[0136] is the gray value of such a pixel). The variogram is calculated as follows: - one for each given direction (i.e. one variogram for each pixel in S and each direction in I). If a variogram is calculated for each given direction, h(O) is the grey level of the initial given pixel before the movement, h(r) is the grey level of the given new pixel after moving a distance r from the initial given pixel along one of the given directions, and the variogram is calculated according to the formula V i (r) = [h(r) - h(O)] 2 where i∈I. Or, - for all given directions simultaneously (i.e. one variogram for each pixel of S). When variograms are calculated simultaneously for all given directions, the step of calculating the variogram of the gray level as a function of distance r is done by averaging the squared differences of h for several pairs of pixels, each at distance r, and the formula is:

[0137]

number

[0138] V Pi (r) is the sum of all pixels P i ∈S. V Pi The formula for each P i ∈S. This particular range of values ​​of r corresponds to the variogram V Pi For evaluation of all necessary parameters of

[0139]

number

[0140] is chosen to allow for convergence. The range R0 is determined depending on the skeletal location and image resolution. In DXA systems, the value of R0 is between 1 cm and 2 cm. The range of the calculations, R0, should not be confused with the range parameter c of the variogram model.

[0141] e) For every pixel P i ∈S for V Pi (r) and tracing or calculating or determining the associated curve on a log-log scale. Variogram curve V on log-log scale Pi The representation means that the values ​​along each axis no longer have units.

[0142] f) V Pi Evaluate the full model V as a least squares regression model of the representation.

[0143] g) Evaluate the parameters of this model including, but not limited to, the initial slope a, sill b, range c, nugget d, and / or area under the curve e. For each pixel's variogram VPi and / or the global variogram V of the sampling S combining the variograms of each pixel, evaluate at least one of the following parameters in a log-log scale: - the initial slope a of the variogram (see 54 in Fig. 8), - sil b, which represents the asymptotic value of the variogram (see 55 in Figure 8), - the range c (see 56 in Figure 8), which represents the distance at which the variogram curve transitions from a quasi-linear progression to an asymptotic behavior; - the nugget d (see 57 in Figure 8), which represents the initial value of the variogram, and - the area under the variogram curve e of the variogram (see 58 in Figure 8). V Pi Or a log-log plot of V is fitted to a mathematical model. These parameters (i.e., coefficients) a, b, c, d, and e of the model are combined to create a bone texture score, B. Each parameter a, b, c, d, and / or e is estimated from a least-squares regression model of the variogram considered. Depending on the content of Image 9, the selected coefficients of the model may vary, as they may not be well-defined (e.g., the variogram curves may not converge to an asymptote and therefore the "range" may not be defined). Preferably, the training method includes evaluating at least one of the parameters sil, b, range, c, nugget, d, and area, e, on a log-log scale, for each variogram of each pixel and / or a global variogram of the sampling S combining the variograms of each pixel. Preferably, the training method includes evaluating at least two of the parameters slope a, sill b, range c, nugget d, and area e on a log-log scale for each variogram for each pixel and / or a global variogram of a sampling S combining the variograms for each pixel. Preferably, the training method comprises evaluating at least one of the following parameters, sill b, extent c, nugget d, area e, with initial slope a and log-log scale, for each variogram of each pixel and / or a global variogram of the sampling S combining the variograms of each pixel: Preferably, the training method comprises evaluating all parameters slope a, sill b, range c, nugget d, area e on a log-log scale for each variogram for each pixel and / or a global variogram of the sampling S combining the variograms for each pixel.

[0144] h) Combining at least one parameter a, b, c, d and / or e into a TBS (dimensionless), e.g. using a linear or non-linear equation depending on the clinical context. For example, the learning method may include the following steps. - combining at least one evaluated parameter a, b, c, d and / or e obtained for the variogram of the pixel or of each pixel into a TBS for the pixel or of each pixel, and / or - combining at least one evaluated parameter a, b, c, d and / or e obtained for the global variogram of the sampling S into a TBS of the sampling S, and / or - combining at least one evaluated parameter a, b, c, d and / or e obtained for the variogram of each pixel into a TBS of the sampling S; Preferably, at least two parameters a, b, c, d, e are combined into the TBS using a linear or non-linear equation depending on the clinical context. In practice, the parameters of the variogram model are combined together into the TBS using a combination equation. By way of example, such a combination equation can include, but is not limited to, a multiple regression model for a given clinical context and anatomical site. In such a case scenario, the TBS is defined as follows: TBS = αa + βb + γc + δd + εe, where the coefficients α, β, γ, δ, and ε are related to the slope, sill, nugget, and area under the curve, respectively. These coefficients are carefully defined prior to the clinical performance optimization stage. These coefficients are typically obtained from different experimental analyses. The best coefficient selection is obtained, for example, on a large set of clinical studies and images, followed by a grid search optimization step. More generally, α, β, γ, δ and / or ε can be obtained in different ways, including: - A table of correction factors based on experimental measurements, and / or - mathematical models and / or simulations of the corrections based on theory, and / or - Machine learning or artificial intelligence (AI) algorithms (where this machine learning or AI is not applied to the training images 9); and / or - Selection of the best coefficients obtained in a grid search optimization phase, for example on a large set of clinical studies and images, with further clinical performance results, and / or - A combination of the above methods; - etc. for example, TBS=a That is, α=1 and β=γ=δ=ε=0. Preferably, the TBS is calculated from a global variogram of a sampling S calculated simultaneously for all given directions.

[0145] i) Optionally, applying robustness improvement steps to the TBS related to at least one patient factor and / or at least one technical factor, preferably applying robustness improvement steps related to both patient and technical factors. The robustness improvement step can be implemented as follows. - before or during the previous step h) (by determining and / or correcting R0, α, β, γ, δ, ε, a, b, c, d and / or e as a function of patient and / or technical factors before or during the determination or calculation of the score B), and / or - after the previous step h) (by correcting score B as a function of patient and / or technical factors). The robustness improvement step may be implemented in different ways (by determining and / or correcting R0, α, β, γ, δ, ε, a, b, c, d, e, and / or b), including the following: - a correction factor table based on experimental measurements that takes into account patient and / or technical factors and their observed effects on the variogram and / or TBS, and / or - mathematical models and / or simulations of theory-based corrections taking into account patient and / or technical factors and their predicted effects on the variogram and / or TBS, and / or - a machine learning or artificial intelligence (AI) algorithm that learns to minimize the effect of patient and / or technical factors on the variogram and / or texture score B (this machine learning or AI is not applied to the training images 9); and / or - selection of the best coefficients obtained in a grid search optimization phase with clinical performance results, for example in a large set of clinical studies and images, up to the point where further patient and / or technical factors have a minimal effect on the variogram and / or TBS, and / or - A combination of the above methods. For example, R0 may be determined and / or corrected as a function of the image resolution of the x-ray acquired image. The at least one patient factor includes: - Effects of the patient's morphology, including at least one of the following: -- the effect of the patient's soft tissue, and / or tissue thickness, and / or its distribution, and / or its composition, and / or indirect surrogates, and / or -- the patient's weight and / or body mass index (BMI) and / or waist circumference, and / or -- The size of the patient, and / or - the effect of at least one pathology or condition of the patient (arthropathy, ascites, aortic calcification, gas, etc.); and / or - The effect of patient positioning during image acquisition. The at least one technical factor includes: - Potentially defective detectors and sensors for acquiring images, and / or - the influence of the scanning mode and settings for acquiring the image, and / or - the technical characteristics of the imaging device used to acquire the images; - the effect of variability between imaging systems for acquiring images, and / or - The signal-to-noise ratio (SNR) of the image, and / or - Image resolution. The robustness improvement step allows: - to be less influenced by the physiological characteristics of the patient, e.g. the volume and / or nature of the soft tissue surrounding the bone; and / or - Less sensitive to the choice of technical parameters of image acquisition. Typically, - The coefficient α is -- Affects reproducibility -- Optimizing fracture prediction - The coefficient β is -- Affects the overall mineralization contribution of (bone) tissue - The coefficient γ is -- Influence the contribution of the overall geometric properties of the (bone) tissue - The coefficient δ is -- related to the signal-to-noise ratio of an image - The coefficient ε is -- Affects reproducibility.

[0146] The first and second artificial intelligences are trained using the same database of training images of the first type 9 and training images of the second type 19. This qualified dataset is necessary for training the deep learning models ANN1 and ANN2, as they rely on supervised learning.

[0147] Such datasets are used for refining, training and validating artificial neural networks (ANN1 and ANN2).

[0148] Each element of the training dataset consists of an X-ray digital radiograph associated with a particular ground truth.

[0149] The ground truth 1, 2 is obtained from bone density and bone texture parameters extracted from DXA scans. This is also possible with other techniques such as (but not limited to) (p)QCT, CT, QUS images. In these scans, BMD T values ​​and bone texture (e.g. TBS) values ​​are obtained. For a given patient, DXA scans from multiple anatomical sites (spine, femur, forearm) are used. The smallest BMD T value is selected as the most relevant for the fracture risk profile. The BMD T values ​​14 and bone texture (e.g. TBS) 15 values ​​are compared with their respective classification thresholds (as shown in Figure 2, these thresholds or categories are “normal”, “osteopenic”, “osteoporotic” for BMD and “normal”, “partially deteriorated”, “deteriorated” for TBS). The resulting stratifications of each score are combined (see Figure 2) to generate a fracture risk category (also called a global score 16). This fracture risk category is labeled with a number from 1 to 9 (or less if other types of categories are defined).

[0150] Depending on the ANN to be trained, the ground truth 1 or 2 data consist of either fracture risk category labels (ANN1 - multi-class classification tasks) or continuous values ​​of BMD T-values ​​and bone texture (e.g. TBS) (ANN2 - regression tasks).

[0151] Matching of ground truth data and X-ray digital radiographs is ensured using anonymized patient identifiers (PIDs). DXA scans9 and X-ray digital radiographs19 are not necessarily obtained on the same day, but ensure that the time lapse between X-ray scans19 and DXA scans9 is short enough (i.e., less than 6 months) to minimize changes in bone status between both modalities.

[0152] The ground truth from a DXA scan9 and its associated X-ray digital radiograph19 do not necessarily originate from the same anatomical site. For simplicity, in what follows, we use the terms DXA and TBS, although other imaging device types and bone texture or structural parameters may be involved.

[0153] Therefore, method 100 is based on a different approach: several artificial neural network (ANN) models ANN1 and ANN2 are defined and trained separately and then combined in one ensemble model to assess the final bone risk category (high risk vs. low risk, as defined in the ground truth of method 100).

[0154] One ANN (ANN1) is designed and trained as a multi-class classifier that takes as input a digital x-ray image 6 to predict a risk category class 16 (classes 1 to 9 reflecting TBS and BMD DXA measurements 14, 15). The other ANN (ANN2) is designed to take the same digital x-ray image 6 as input and predict a set of continuous values ​​of BMD T-values ​​14 and raw TBS 15.

[0155] The use of both models allows a final bone risk assessment, which is evaluated in a second stage using a dedicated ensemble model, based on the ANN predictions (categorical and continuous values) and on the metadata values.

[0156] The optimization of these models is ensured during a dedicated training phase with the final goal of minimizing the false positive rate and avoiding overfitting. In fact, the models (ANN1 and ANN2) are trained using data augmentation and cross-validated to obtain maximum predictive performance without bias induced during training.

[0157] In this first model approach of ANN1, one deep artificial neural network ANN1 (e.g., including but not limited to, a convolutional neural network (CNN) or a recurrent neural network (RNN)) is implemented and trained to predict the “fracture risk profile” class. During the training phase, ANN1 receives as input the preprocessed digital bone X-ray based images19 and their corresponding labels. A dedicated preprocessing of the X-ray input images is performed in the same way for both training and evaluation. The preprocessing ensures the automation of bone ROI selection and its match with the input size of the model of the method100, while preserving as much resolution as possible. The principle of this deep neural network ANN1 is to benefit from a strong backbone architecture and high resolution X-ray images to extract the most important feature map information that allows a correct ground truth risk profile classification. The output of this classifier consists of a vector of size 9, the index of whose maximum value is used as the predicted class.

[0158] The second model approach of ANN2 is a deep ANN that infers the continuous values ​​of minimum BMD T value and raw TBS. It can be presented as a deep regression model that outputs a set of two continuous values ​​obtained from the regression output layer. The backbone architecture reduces the input X-ray images to a high DXA-like resolution, on which feature extraction enables the regression and computation tasks of minimum BMD T value and raw TBS.

[0159] The pre-processing stage 17 (of the pre-processing, testing, or validation phase) modifies or labels the X-ray digital images so that they can be fed to the artificial neural network for inference (in both the training and evaluation phases). Pre-processing includes, but is not limited to, the following steps: Labeling of potential artifacts (e.g. metals, implants) · Uniform appearance of X-ray radiographs (e.g. resolution, CR / BR, size) ·ROI focus / optimization Data Augmentation (learning methods only) Identifying anatomical locations ·segmentation Grayscale pixel values -Normalize pixel values ​​to [0,1] Quality assessment of radiographic imaging

[0160] Method 100 also includes the following steps prior to analysis steps 11, 12, and 13 (during the testing and clinical optimization phases of FIG. 6). - constructing a third training set by performing the following steps multiple times: --obtaining training images 9 of a first type (of a third training set) showing the cancellous bone portion of the training bones; -- acquiring training images 19 of a second type (of a third training set) that are X-ray based images showing the same training bone, but not necessarily showing the cancellous bone portion thereof; -- Steps to get metadata 18 -- determining from the training images 9 of the first type of the third training set by technical means (but without implementing artificial intelligence on the images 9): --- a density score 14 that depends on or is equal to the bone density of the training bones shown in the first type training images of the third training set. Typically, the bone density is determined directly from the first type training images as DXA metadata of the first type training images; and --- a Trabecular Bone Score (TBS) 15 that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type 9 of the third training set. Typically, this TBS 15 is determined by calculating or determining an empirical variogram of the grey levels of the training images of the first type (US Pat. No. 5,399,433) (see also steps a) to i) above). -- determining, by technical means (but without implementing artificial intelligence on the image 9), a global score 16 that depends on this density score and the cancellous bone score, from: --- a density score 14 that depends on the bone density of the training bones shown in the training images 9 of the first type of the third training set; and --- A texture-dependent trabecular bone score (TBS) 15 of the trabecular bone portion of the training bone shown in the training image 9 of the first type of the third training set - implementing first and second analyses 11, 12 by the first and second artificial intelligences on training images of a second type of a third training set, and training a third artificial intelligence by learning from: the difference between the scores (density score 14, TBS 15 and / or global score 16) obtained from a training image of the first type 9 of the third training set without ANN1 and ANN2 and the scores (density score 14, TBS 15 and / or global score 16) obtained from a training image of the second type 19 of the same training bone of the third training set using ANN1 and ANN2; and -- Metadata 18.

[0161] In a preferred embodiment of the method 100, - AI3 is a classification and regression tree (CART) to evaluate whether to flag with an associated confidence score. - This training step is done by learning from the DXA ground truth (flagged or unflagged) via supervised learning. - This training is done using a grid search over the tree architecture (optimizing vertical depth, number of terminal nodes, maximum features to consider for splitting nodes, etc.) to optimize the accuracy score until the minimum test loss is reached, while avoiding overfitting.

[0162] The first training set and the second training set may be the same training set.

[0163] The third training set is not the same training set as the first training set and / or the second training set, because the third training set (used in the clinical optimization phase) is used to optimize the method 100 after training ANN1 and ANN2 based on the first training set and / or the second training set.

[0164] The first artificial intelligence, the second artificial intelligence, and the third artificial intelligence are trained separately.

[0165] The validation phase (left part of Fig. 7) consists of the steps described above for fully optimizing all ANNs that are relevant for clinical aspects. i (i=1 or 2) is important to prepare and validate the models. At this stage, the ANN1 and ANN2 models are tested in external cohorts to confirm their robustness and ability to predict the final clinical outcome.

[0166] In this validation phase, all models, their configurations, and their parameters are sealed, time-stamped, tracked, and packaged. The mechanism used to ensure traceability and tracking of parameters is like a blockchain. All parameters, models, metadata, code versions, metrics, tags, labels, output files, and results are logged in a time-stamped record. A step included in the validation phase is to ensure persistence of all configurations of ANNi models as records.

[0167] All these operational steps at this stage are in line with and comply with regulatory requirements in the field of software as a medical device incorporating AI and ANN technologies, and follow best technology practices in this field.

[0168] At the end of this stage, all ANNi models will be successfully validated and ready for integration into the final product during the industrialization stage.

[0169] Once all ANNi models have been verified in the validation phase, the final global module is integrated within the final product in the industrialization phase (right part of Figures 1 and 7).

[0170] The traceability of all ANNi learning experiments is ensured at the previous stage from ground truth to deployment. This traceability principle guarantees the suitability of the model with optimized parameters for clinical validation to the final product. It also allows the right deployed module version in the field, together with the possibility to update or upgrade this right module when better optimization is performed in the offline continuous improvement flow.

[0171] This module is dynamically loaded into the product to provide new functionality as a service in accordance with the teachings of this invention.

[0172] The apparatus according to the invention comprises technical means (in particular means arranged and / or programmed and / or configured for, respectively, calculating, determining, obtaining, selecting, calculating, evaluating, combining, applying an improvement step, receiving images, implementing artificial intelligence, implementing analysis, presenting the results, constructing a learning set, training the artificial intelligence) arranged and / or programmed and / or configured to implement all the aforementioned steps (in particular, the steps of calculating, determining, obtaining, selecting, calculating, evaluating, combining, applying an improvement step, receiving images, implementing artificial intelligence, implementing analysis, presenting the results, constructing a learning set, training the artificial intelligence).

[0173] Typically, at least one of the aforementioned means of the device according to the invention, preferably each means of the device according to the invention (in particular means arranged and / or programmed and / or configured for calculating, determining, obtaining, selecting, calculating, evaluating, combining, applying an improvement step, receiving images, implementing artificial intelligence, implementing an analysis, presenting results, building a training set, training artificial intelligence) is a technical means.

[0174] Typically, each means of the device according to the invention implementing the aforementioned steps (in particular means arranged and / or programmed and / or configured for calculating, determining, obtaining, selecting, calculating, evaluating, combining, applying improvement steps, receiving images, implementing artificial intelligence, implementing analysis, presenting results, building a training set, training artificial intelligence) comprises at least one computer, one central processing unit or computing unit, one analog electronic circuit (preferably dedicated), one digital electronic circuit (preferably dedicated), and / or one microprocessor (preferably dedicated), and / or software means.

[0175] The device according to the invention comprises: - means for implementing the steps of acquiring images 6 and / or 9 and / or 19. These means for acquiring digitized images typically comprise: -- Conventional X-ray imaging systems, and / or -- Digital X-ray imaging systems, and / or -- Dual-energy X-ray absorptiometry (DXA) imaging systems, and / or -- Projection computed tomography (CT) imaging systems, and / or -- Quantitative Computed Tomography (QCT) imaging systems, and / or -- projection quantitative computed tomography imaging systems, and / or -- a peripheral quantitative computed tomography (pQCT) imaging system, and / or -- a high-resolution peripheral quantitative computed tomography (HR-pQCT) imaging system, and / or -- combinations thereof, and - means for implementing the aforementioned steps (in particular the steps of calculating, determining, obtaining, selecting, calculating, evaluating, combining, applying an improvement step, receiving images, implementing artificial intelligence, implementing analysis, presenting results, building a training set, and training an artificial intelligence), which are typically combined in a single computer; - A screen arranged to display the output 16 of ANN1 and / or the outputs 14 and / or 15 of ANN2 and / or the conclusions or outputs of AI3.

[0176] This embodiment also includes the following: - a computer program comprising instructions which, when executed by a computer, implement the steps of the method 100; and / or - a computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method 100; and / or - a computer-readable storage medium containing instructions that, when executed by a computer, cause the computer to perform the steps of method 100.

[0177] Of course, the invention is not limited to the examples described here, to which numerous modifications can be made without going beyond the scope of the invention.

[0178] For example, a variation of the method 100 (and the corresponding apparatus) could include only ANN1 (without ANN2 and AI3) or only ANN2 (without ANN1 and AI3), in which the method 100 (described only for its differences compared to the above description of Figures 1 to 8) is a method for analyzing bone texture from a digital image 6 obtained by imaging selected areas containing bone structures, and includes: - receiving an input X-ray image 6 showing an input bone, - a step of bone score analysis (respectively 11 or 12 above) of the received input X-ray image 6 by a bone score artificial intelligence (ANN1 or ANN2 respectively) implemented by technical means, which gives as a result of this bone score analysis: -- a global score 16 (i.e., when the bone score artificial intelligence is ANN1), which depends at least on a trabecular bone score (TBS), which depends on the texture of the trabecular bone portion of the input bone shown in the received input X-ray image; and / or -- a density score 14 that depends on (or consists of) the bone density of the input bone shown in the received input X-ray image, and / or a trabecular bone score (TBS) 15 that depends on the texture of the trabecular bone portion of the input bone shown in the received input X-ray image (i.e., when the bone score artificial intelligence is ANN2).

[0179] The bone score artificial intelligence ANN1 or ANN2 is a neural network.

[0180] Further in this variation, if the bone scoring artificial intelligence is ANN1, the method 100 includes, prior to the analysis step 11, the following training as described above for ANN1. - Construct a first training set by performing the following steps multiple times: - obtaining a first type of training image 9 showing the cancellous part of the training bone; --acquiring a related second type of training image 12, which is an X-ray based image, showing the same training bone, but not necessarily showing the cancellous bone portion thereof; -- determining by technical means (but without implementing artificial intelligence on the images 9) from training images 9 of a first type: --- a density score 14 that depends on the bone density of the training bones shown in the training images of the first type, and --- Trabecular bone score (TBS) 15, which depends on the texture of the trabecular bone portion of the training bone shown in the first type of training image -- determining by technical means (but without implementing artificial intelligence on the images 9 and / or the pre-determined density scores 14 and / or the pre-determined TBS 15) a global score 16 that depends on these density scores and the cancellous bone score from: --- a density score 14 that depends on the bone density of the training bones shown in the training images of the first type, and --- Trabecular bone score (TBS) 15, which depends on the texture of the trabecular bone portion of the training bone shown in the first type of training image - training the bone score artificial intelligence ANN1 by providing the bone score artificial intelligence with a second type of training image 19 together with its associated ground truth including or consisting of a global score 16 determined for the first type of training image 9 associated with this second type of training image 19.

[0181] Further in this variation, if the bone scoring artificial intelligence is ANN2, the method 100 includes, prior to the analysis step 12, the following training as described above for ANN2. - Construct a second training set by performing the following steps multiple times: - obtaining a first type of training image 9 showing the cancellous part of the training bone; --acquiring a related second type of training image 19, which is an X-ray based image, showing the same training bone, but not necessarily showing the cancellous bone portion thereof. -- determining by technical means (but without implementing artificial intelligence on the images 9) from training images of a first type: --- a density score 14 depending on the bone density of the training bone shown in the training image of the first type, and / or --- Trabecular bone score (TBS) 15, which depends on the texture of the trabecular bone portion of the training bone shown in the first type of training image - training the bone score artificial intelligence by providing the second type of training images to the bone score artificial intelligence along with a ground truth that includes or consists of a density score determined for the first type of training images 9 associated with this second type of training images 19 and / or a trabecular bone score determined for the first type of training images 9 associated with this second type of training images 19.

[0182] In another variation, the method 100 (and corresponding apparatus) may include ANN1 and ANN2 without AI3.

[0183] Of course, different features, aspects, modifications and embodiments of the invention can be combined with each other in various combinations, unless they are incompatible or mutually exclusive.

Claims

1. 1. A method for analyzing bone texture, comprising: receiving an input digitized X-ray image (6) showing an input bone; a step of bone score analysis (11, 12) of said received input X-ray image by a bone score artificial intelligence implemented by technical means, said bone score artificial intelligence giving as a result of said bone score analysis, without calculation or determination of an empirical variogram of the grey levels of said received input X-ray image (6), a global score (16) depending at least on a trabecular bone score (TBS) quantifying local variations in grey levels from the empirical variogram of grey levels of the trabecular bone portion of the input bone shown in the received input X-ray image, and / or -- a Trabecular Bone Score (TBS) (15) quantifying the local variation in grey levels from the empirical variogram of grey levels of the trabecular bone portion of the input bone shown in the received input X-ray image; A method for analyzing bone texture, including:

2. The Bone Score Artificial Intelligence (ANN １ , ANN ２ 2. The method of claim 1, wherein the network is a neural network.

3. Constructing a training set by performing the following steps multiple times: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; determining, by technical means, from said training images of the first type: a density score (14) depending on the bone density of the training bones shown in the training images of the first type; and a trabecular bone score (TBS) (15) depending on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type; and determining from said trabecular bone scores (TBS) dependent on the texture of the trabecular bone portions of said training bones shown in said training images of said first type a global score (16) dependent on these density scores and on the trabecular bone scores; The second type of training image (19) is then subjected to a bone score artificial intelligence (ANN) process together with its associated ground truth including the global score determined for the first type of training image (9) associated with the second type of training image (19). １ ANN １ ) learning process; The method of claim 1 or claim 2, comprising:

4. Constructing a training set by performing the following steps multiple times: -- obtaining training images of a first type (9) showing the cancellous portion of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means, a density score (14) depending on the bone density of the training bones shown in the training images of the first type, and / or determining from said first type of training image a trabecular bone score (TBS) (15) depending on the texture of the trabecular bone portion of said training bone shown in said first type of training image: the density score determined for the training image (9) of the first type associated with this training image (19) of the second type, and / or - including the cancellous bone score determined for the training image (9) of the first type associated with this training image (19) of the second type, and inputting the training image of the second type together with its associated ground truth to the bone score artificial intelligence (ANN) ２ ) to train the bone score artificial intelligence; The method of claim 1 or claim 2, comprising:

5. receiving said input X-ray image (6) showing an input bone; The first artificial intelligence (ANN) implemented by technical means １ a first analysis (11) of the received input X-ray image by the first artificial intelligence, -- a density score depending on the bone density of the input bone shown in the received input X-ray image; -- a trabecular bone score (TBS) depending on the texture of the trabecular bone portion of the input bone shown in the received input X-ray image. giving a global score (16) which depends on both Second artificial intelligence implemented by technical means (ANN ２ a second analysis (12) of the received input X-ray image by the second artificial intelligence, a density score (14) depending on the bone density of the input bone shown in the received input X-ray image, and / or -- providing said trabecular bone score (TBS) (15) dependent on the texture of the trabecular bone portion of said input bone shown in said received input X-ray image; Third artificial intelligence (AI) implemented by technical means ３ a third analysis step by the third artificial intelligence having as input the results of the first and second analyses and having as output a result that depends on the consistency of the result of the first analysis with the result of the second analysis; The method of claim 1 or claim 2, comprising:

6. 6. The method of claim 5, wherein the third artificial intelligence uses as further input (18) at least one of the following: the age of the patient from whom the received input X-ray image was acquired, the gender of the patient from whom the received input X-ray image was acquired, the body type of the patient from whom the received input X-ray image was acquired, the type of equipment from which the received input X-ray image was acquired, and / or acquisition parameters of the equipment from which the received input X-ray image was acquired.

7. The method of claim 5 , wherein the first artificial intelligence is a neural network and the second artificial intelligence is a neural network.

8. The method of claim 5 , wherein the first artificial intelligence, the second artificial intelligence, and the third artificial intelligence are three different artificial intelligences.

9. The method according to claim 5 , wherein the technical means for implementing the first artificial intelligence, the second artificial intelligence and the third artificial intelligence are the same technical means.

10. The first training set is subjected to the following steps: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score dependent on the bone density of the training bones shown in the training images of the first type; and determining a trabecular bone score (TBS) dependent on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type; and determining a global score that depends on the density scores and the cancellous bone scores from the cancellous bone scores (TBS) that depend on the texture of the cancellous bone portions of the training bones shown in the training images of the first type, multiple times to construct a first training set; The training image (19) of the second type is fed to the first artificial intelligence (ANN) together with its associated ground truth including the global score (16) determined for the training image (9) of the first type associated with the training image (19) of the second type. １ ) to train the first artificial intelligence; The method of claim 5 , comprising:

11. The steps of: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score (14) depending on the bone density of the training bones shown in the training images of the first type, and / or determining a trabecular bone score (TBS) (15) that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; multiple times to construct a second training set; the density score (14) determined for the training image (9) of the first type associated with this training image (19) of the second type, and / or - the second type of training image (19) together with its associated ground truth including the cancellous bone score (15) determined for the first type of training image (9) associated with this second type of training image (19) is fed to the second artificial intelligence (ANN) ２ ) to train the second artificial intelligence; The method of claim 5 , comprising:

12. The method of claim 10 , wherein the first artificial intelligence and the second artificial intelligence are trained using the same database of first type training images and second type training images.

13. The steps of: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means from a first type of learning image (9) a density score (14) depending on the bone density of the training bones shown in the training images of the first type; and determining a trabecular bone score (TBS) (15) that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type (9); and determining from said trabecular bone score (TBS) which depends on the texture of the trabecular bone portion of said training bones shown in said first type of training images (9), a global score (16) which depends on this density score and on the trabecular bone score; -- implementing said first and second analyses by said first and second artificial intelligences on said second type of training images (19); constructing a third training set by performing The third artificial intelligence (AI) learns from the difference between the scores (14, 15, 16) obtained from the first type of learning image (9) and the scores (14, 15, 16) obtained from the second type of learning image (19) of the same learning bone. ３ ) learning process; The method of claim 5 , comprising:

14. The method of claim 10 , wherein the first artificial intelligence, the second artificial intelligence, and the third artificial intelligence are trained separately.

15. The method of claim 3 , wherein the first type of training images and the second type of training images are acquired of the same training bone and are acquired less than six months apart.

16. 4. The method of claim 3, wherein the first type of training images are dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

17. 4. The method of claim 3, wherein the second type of training images are not dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

18. 3. The method of claim 1 or claim 2, wherein the received input X-ray image is not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

19. The method of claim 1 or claim 2, wherein the received input x-ray image is a digital x-ray image having a spatial resolution of less than 1 mm per pixel.

20. 1. An apparatus for analyzing bone texture, comprising: - means arranged and / or programmed and / or configured to receive an input digitized X-ray image (6) representative of an input bone; a bone score artificial intelligence (ANN) arranged and / or programmed and / or configured to implement a bone score analysis (11, 12) of said received input X-ray images; １ , ANN ２ ), as a result of this bone score analysis, without calculation or determination of an empirical variogram of the grey levels of the received input X-ray image (6), a global score (16) depending at least on a trabecular bone score (TBS) quantifying the local variations in grey levels from said empirical variogram of grey levels of the trabecular bone portion of said input bone shown in said received input X-ray image, and / or said bone scoring artificial intelligence arranged and / or programmed and / or configured to provide a cancellous bone score (TBS) (15) quantifying said local variations in grey levels from said empirical variogram of grey levels of said cancellous bone portion of said input bone shown in said received input X-ray image; 1. An apparatus for analyzing bone texture comprising:

21. 21. The apparatus of claim 20, wherein the bone scoring artificial intelligence is a neural network.

22. The steps of: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score (14) depending on the bone density of the training bones shown in the training images of the first type; and determining a trabecular bone score (TBS) (15) that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type; and and determining a global score (16) that depends on these density scores and the cancellous bone score (TBS), which depends on the texture of the cancellous bone portion of the training bone shown in the training images of the first type, multiple times to build a training set. The second type of training image (19) is then subjected to a bone score artificial intelligence (ANN) process together with its associated ground truth including the global score determined for the training image (9) of the first type associated with the second type of training image (19). １ ANN １ (a) a means arranged and / or programmed and / or configured to train a 22. The apparatus of claim 20 or claim 21, comprising:

23. The steps of: -- obtaining training images of a first type (9) showing the cancellous part of a training bone; -- acquiring a related second type of training image (19), which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score (14) depending on the bone density of the training bones shown in the training images of the first type, and / or determining a trabecular bone score (TBS) (15) that depends on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; a plurality of times to construct a training set; and the density score determined for the training image (9) of the first type associated with this training image (19) of the second type, and / or - the second type of training image (19) together with its associated ground truth including the cancellous bone score determined for the first type of training image (9) associated with this second type of training image (19) is processed by a bone score artificial intelligence (ANN) ２ means arranged and / or programmed and / or configured to train said bone scoring artificial intelligence by providing a 22. The apparatus of claim 20 or claim 21, comprising:

24. - means arranged and / or programmed and / or configured to receive said input X-ray image (6) showing an input bone; a first artificial intelligence (ANN) arranged and / or programmed and / or configured to implement a first analysis (11) of the received input X-ray image; １ ) , wherein the first analysis results in a density score dependent on the bone density of the input bone shown in the received input X-ray image; and said first artificial intelligence arranged and / or programmed and / or configured to provide a global score (16) that depends on both the texture of the cancellous bone portion of said input bone shown in said received input X-ray image and a cancellous bone score (TBS) that depends on the texture of the cancellous bone portion of said input bone; a second artificial intelligence (ANN) arranged and / or programmed and / or configured to implement a second analysis (12) of the received input X-ray image; ２ ) as a result of said second analysis said density score (14) being dependent on the bone density of said input bone shown in said received input X-ray image, and / or a second artificial intelligence arranged and / or programmed and / or configured to provide said trabecular bone score (TBS) (15) dependent on the texture of the trabecular bone portion of said input bone shown in said received input X-ray image; a third artificial intelligence (AI) arranged and / or programmed and / or configured to implement a third analysis; ３ ) arranged and / or programmed and / or configured to have as input the results of the first and second analyses and to have as output a result that depends on the consistency of the result of the first analysis with the result of the second analysis; 22. The apparatus of claim 20 or claim 21, comprising:

25. 25. The apparatus according to claim 24, wherein the third artificial intelligence is arranged and / or programmed and / or configured to use at least one of the following as further inputs (18): the age of the patient from whom the received input X-ray image was acquired, the sex of the patient from whom the received input X-ray image was acquired, the body type of the patient from whom the received input X-ray image was acquired, the type of equipment from which the received input X-ray image was acquired, and / or acquisition parameters of the equipment from which the received input X-ray image was acquired.

26. 25. The apparatus of claim 24, wherein the first artificial intelligence is a neural network and the second artificial intelligence is a neural network.

27. 25. The apparatus of claim 24, wherein the first artificial intelligence, the second artificial intelligence, and the third artificial intelligence are three different artificial intelligences.

28. The apparatus according to claim 24, wherein the technical means for implementing the first artificial intelligence, the second artificial intelligence and the third artificial intelligence are the same technical means.

29. The steps of: --obtaining a first type of training image showing a cancellous portion of a training bone; --acquiring a related second type of training image, which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score dependent on the bone density of the training bones shown in the training images of the first type; and determining a trabecular bone score (TBS) dependent on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type; and From said cancellous bone scores (TBS) which depend on the texture of the cancellous bone parts of said training bones shown in said training images of said first type, determining a global score which depends on these density scores and cancellous bone scores; a plurality of times to construct a first training set; and means arranged and / or programmed and / or configured to train the first artificial intelligence by providing said second type of training image (9) together with its associated ground truth including said global score determined for said first type of training image (9) associated with said second type of training image (19); 25. The apparatus of claim 24, comprising:

30. The steps of: --obtaining a first type of training image showing a cancellous portion of a training bone; --acquiring a related second type of training image, which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score depending on the bone density of the training bones shown in the training images of the first type; and / or determining a trabecular bone score (TBS) dependent on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; a plurality of times to construct a second training set; and the density score determined for the training image (9) of the first type associated with this training image (19) of the second type, and / or means arranged and / or programmed and / or configured to train said second artificial intelligence by providing said second type of training image together with its associated ground truth comprising said cancellous bone score determined for said training image (9) of said first type associated with said second type of training image (19); 25. The apparatus of claim 24, comprising:

31. 30. The apparatus of claim 29, wherein the first artificial intelligence and the second artificial intelligence are arranged and / or programmed and / or configured to learn using the same database of first type training images and second type training images.

32. The steps of: --obtaining a first type of training image showing a cancellous portion of a training bone; --acquiring a second type of training image, which is an X-ray based image, showing the same training bone; -- by technical means from said first type of training images a density score dependent on the bone density of the training bones shown in the training images of the first type; and determining a trabecular bone score (TBS) dependent on the texture of the trabecular bone portion of the training bones shown in the training images of the first type; -- by technical means, the density score being dependent on the bone density of the training bones shown in the training images of the first type; and determining from said trabecular bone score (TBS) dependent on the texture of said trabecular bone portion of said training bones shown in said training images of said first type a global score dependent on said density score and said trabecular bone score; implementing first and second analyses by first and second artificial intelligences on the second type of training images; a plurality of times to construct a third training set; and means arranged and / or programmed and / or configured to train a third artificial intelligence by learning from the difference between the scores obtained from training images of the first type and scores obtained from training images of the second type of the same training bone; 21. The apparatus of claim 20, comprising:

33. 30. An apparatus as claimed in claim 29, arranged and / or programmed and / or configured to train the first artificial intelligence, the second artificial intelligence and the third artificial intelligence separately.

34. The apparatus described in claim 22, wherein the means arranged and / or programmed and / or configured to train a first artificial intelligence and the means arranged and / or programmed and / or configured to train a second artificial intelligence are arranged and / or programmed and / or configured together to check that the first type of training image and the second type of training image are obtained of the same training bone and are obtained less than six months apart.

35. 23. The apparatus of claim 22, wherein the first type of training images are dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

36. 23. The apparatus of claim 22, wherein the second type of training images are not dual-energy X-ray absorptiometry (DXA) images, peripheral quantitative computed tomography ((p)QCT) images and / or high-resolution peripheral quantitative computed tomography (HR-pQCT) images, computed tomography (CT) images, or quantitative ultrasound (QUS) images.

37. 22. The apparatus of claim 20 or claim 21, wherein the received input X-ray image is not a dual-energy X-ray absorptiometry (DXA) image, a peripheral quantitative computed tomography ((p)QCT) image and / or a high-resolution peripheral quantitative computed tomography (HR-pQCT) image, a computed tomography (CT) image, or a quantitative ultrasound (QUS) image.

38. 22. Apparatus according to claim 20 or claim 21, wherein the received input X-ray image is a digital X-ray image having a spatial resolution of less than 1 mm per pixel.

39. A computer program comprising instructions for performing the steps of the method according to any one of claims 1 or 2 when executed by a computer.

40. A computer-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method according to claim 1 or claim 2.

Images