Publicação
Radiographic annotations of the canine hip joint conformation for computer software training in the diagnosis of dysplasia
| Resumo: | This dissertation begins with two reviews, one on canine hip dysplasia (CHD) and another on artificial intelligence in veterinary imaging; then two studies are presented, one on hip distraction index (DI) computation and another on femoral neck thickness index (FNTi). CHD is a hereditary orthopedic disease that predominates in large and giant breeds, causing lameness and disability. To assess the presence of this pathology in young dogs, hip laxity is evaluated on distraction-stress radiographs, using a hip distractor and calculating the DI. In adult dogs the evaluation is performed on ventrodorsal hip-extended radiographs, examining for evidence of congruity changes and osteoarthritis. Artificial intelligence is emerging in the field of veterinary imaging. Medical images are very subjective to evaluate, therefore the application of artificial intelligence and deep learning methods to automate the analysis process would be a great resource. For the DI computation, first a newly trained examiner (TEx) was tested by comparing the DI measurements with an experienced examiner (EEx) on 50 images (100 hips). Then, the TEx performed annotations on 174 images (348 hips) to train the AI model for the DI computation. The 50 images (100 hips) mentioned previously were used to test the model. In the comparison between EEx and TEx DI measurements, the intra-class correlation (ICC) test and the Bland-Altman plot denoted adequate reproducibility. The comparison between the TEx and the AI model also showed agreement and reliability. The Bland-Altman analysis revealed significant mean differences (p < 0.001) between the TEx and AI model, however, the 95% LoA can be interpreted as clinically small, because 95% of all measurement differences lie in a short range. Additionally, the ICC values showed excellent reliability. This could mean that the measurements are related and functionally consistent. Nevertheless, the AI model would benefit from a larger image dataset for training. The FNTi research began by using previously trained AI models that determined femoral head diameter for other CHD research purposes in the Dys4vet project. Simultaneously, specific annotations in the femoral neck in 25 images were performed. The aim was to develop an algorithm to teach the AI software to determine the minimum FNT, in order to compute the FNTi. Following this, 29 images (58 hips) were randomly selected with all five FCI categories similarly represented. All hips were annotated by an examiner in 2 sessions to evaluate repeatability and afterwards, the measurements from the examiner were compared against the AI algorithm, to evaluate its performance. In the end, the association of the FNTi with the degree of CHD assessed by the FCI score was studied. The results showed adequate intraexaminer as well as between the examiner and the AI algorithm agreement and reliability. Mean FNTi values were shown to be gradually higher in the different FCI categories (p < 0.05). In conclusion a larger dataset would be beneficial for both scientific works. Nevertheless, the automated DI calculation would definitely be a great asset to veterinarians, as it would be less reliant on expert knowledge. The new FNTi parameter has the potential to make the classification of CHD more objective if it was included as a scoring criterion and it is promising in terms of its application through a computer vision model capable of autonomously identifying and classifying CHD. |
|---|---|
| Autores principais: | Pereira, Ana Inês Figueiredo Bernardes |
| Assunto: | Hip Dysplasia screening Imaging |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Trás-os-Montes e Alto Douro |
| Idioma: | inglês |
| Origem: | Repositório da UTAD |
| Resumo: | This dissertation begins with two reviews, one on canine hip dysplasia (CHD) and another on artificial intelligence in veterinary imaging; then two studies are presented, one on hip distraction index (DI) computation and another on femoral neck thickness index (FNTi). CHD is a hereditary orthopedic disease that predominates in large and giant breeds, causing lameness and disability. To assess the presence of this pathology in young dogs, hip laxity is evaluated on distraction-stress radiographs, using a hip distractor and calculating the DI. In adult dogs the evaluation is performed on ventrodorsal hip-extended radiographs, examining for evidence of congruity changes and osteoarthritis. Artificial intelligence is emerging in the field of veterinary imaging. Medical images are very subjective to evaluate, therefore the application of artificial intelligence and deep learning methods to automate the analysis process would be a great resource. For the DI computation, first a newly trained examiner (TEx) was tested by comparing the DI measurements with an experienced examiner (EEx) on 50 images (100 hips). Then, the TEx performed annotations on 174 images (348 hips) to train the AI model for the DI computation. The 50 images (100 hips) mentioned previously were used to test the model. In the comparison between EEx and TEx DI measurements, the intra-class correlation (ICC) test and the Bland-Altman plot denoted adequate reproducibility. The comparison between the TEx and the AI model also showed agreement and reliability. The Bland-Altman analysis revealed significant mean differences (p < 0.001) between the TEx and AI model, however, the 95% LoA can be interpreted as clinically small, because 95% of all measurement differences lie in a short range. Additionally, the ICC values showed excellent reliability. This could mean that the measurements are related and functionally consistent. Nevertheless, the AI model would benefit from a larger image dataset for training. The FNTi research began by using previously trained AI models that determined femoral head diameter for other CHD research purposes in the Dys4vet project. Simultaneously, specific annotations in the femoral neck in 25 images were performed. The aim was to develop an algorithm to teach the AI software to determine the minimum FNT, in order to compute the FNTi. Following this, 29 images (58 hips) were randomly selected with all five FCI categories similarly represented. All hips were annotated by an examiner in 2 sessions to evaluate repeatability and afterwards, the measurements from the examiner were compared against the AI algorithm, to evaluate its performance. In the end, the association of the FNTi with the degree of CHD assessed by the FCI score was studied. The results showed adequate intraexaminer as well as between the examiner and the AI algorithm agreement and reliability. Mean FNTi values were shown to be gradually higher in the different FCI categories (p < 0.05). In conclusion a larger dataset would be beneficial for both scientific works. Nevertheless, the automated DI calculation would definitely be a great asset to veterinarians, as it would be less reliant on expert knowledge. The new FNTi parameter has the potential to make the classification of CHD more objective if it was included as a scoring criterion and it is promising in terms of its application through a computer vision model capable of autonomously identifying and classifying CHD. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.