Two steps closer to automated canine hip dysplasia diagnosis

• Autor(a) principal: Silva, Diogo Emanuel Moreira da
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: https://hdl.handle.net/10348/13336
• topic: canine hip dysplasia; machine learning; deep learning; segmentation; medical images

Canine hip dysplasia is a hereditary disease that progressively leads to instability and laxity of the hip joint, which is to say a poor congruence between the femoral head and the acetabulum. This disease mainly affects large dogs and represents 29% of all orthopaedic cases. This factor, combined with the high debilitating consequences of this disease, demonstrates the need for a new solution to prevent these cases. Since this disease has no cure, selective breeding is the only way to reduce its incidence. Therefore, the earlier the diagnosis, the greater the likelihood of preventing breeding and avoiding the genetic transmission of this condition. Currently, the diagnosis of this disease is performed through the visual analysis of radiographic images by specialists. This diagnosis is intrinsically subjective, depending on the scrutiny and experience of the specialist. Recently, the emergence of Deep Learning has made it possible to build support systems for medical diagnosis, reaching performance levels comparable to or superior to human specialists. The present dissertation intends to lay foundations, in two aspects, for the construction of software for the automated diagnosis of this disease. First, Deep Learning algorithms were developed for femur and acetabulum segmentation. The correct segmentation of these two bone structures allows an objective and quantifiable measure of hip joint congruence, an important metric for the final diagnosis. Nevertheless, the construction of these algorithms has the premise of needing a large volume of images annotated by specialists. Furthermore, semantic annotation of the femur and acetabulum is a very time-consuming and rigorous task. As such, in a second moment, Active Learning techniques were explored. These techniques aim to efficiently select which unannotated images will benefit the model the most. In other words, these techniques aim to maximise the performance of a model with the smallest possible number of annotated images.The segmentation models’ results are promising, reaching state-of-the-art performance with a dice score of 0.981, surpassing a literature model dice score of 0.970, and more significant gains in harder-to-segment classes. The Active Learning results reveal that a proposed method leads to an average 10.8% performance increase over a baseline and a possible reduction of almost 19% of data labelled by veterinary professionals, thus reducing their time spent in the annotation process.