In vivo Automatic Cataract Detection and Classification using the Ultrasound Technique

Abstract

A catarata é uma condição associada à perda da normal transparência do cristalino. A sua progressão pode resultar na perda parcial ou total da visão. O processo de formação da catarata resulta no aumento do número de dispersores de luz no cristalino, na sua opacificação e no aumento da sua dureza. Atualmente, a catarata não tem prevenção ou tratamento farmacológico, pelo que a única terapêutica eficaz para recuperação da visão é a remoção cirúrgica da catarata com substituição do cristalino por uma lente intraocular. A facoemulsificação é a técnica cirúrgica mais usada. A detecção precoce da catarata, a classificação da sua severidade, bem como a estimação da sua dureza, constituem aspetos clínicos relevantes. A estimação da dureza da catarata é importante para a seleção adequada dos níveis de energia de facoemulsificação e consequente diminuição das complicações cirúrgicas. Este trabalho tinha por objetivo a deteção precoce e caraterização da catarata, nomeadamente a estimação da sua dureza, recorrendo a métodos por ultrassons, bem como a sua classificação automática usando técnicas de “machine learning”. De modo a demonstrar a aplicabilidade e utilidade dos ultrassons na caracterização não invasiva da catarata, foi realizado, inicialmente, um estudo ex vivo de prova de conceito em olhos de suíno, induzindo nestes catarata por imersão dos cristalinos numa solução de etanol:2.propanol:formalina. A aquisição dos sinais acústicos foi realizada com um transdutor ultrassónico de 25 MHz, funcionando em modo pulso-eco, tendo sido extraídos os parâmetros velocidade e atenuação. De modo a visualizar e caracterizar a catarata, foram construídas imagens B-scan e paramétricas de Nakagami, com base nos sinais de backscattering. Foram extraídas 97 features, para a classificação automática dos graus de severidade da catarata, usando os classificadores de Bayes, K-Nearest-Neighbours, Fisher Linear Discriminant e Support Vector Machine (SVM). A distribuição espacial dos agregados proteicos e a compactação de fibras para os diferentes graus de severidade da catarata foram analisadas com base em imagens histológicas dos cristalinos obtidas nas regiões do córtex e do núcleo. O estudo ex vivo demonstrou que a dureza da catarata pode ser estimada de forma não invasiva e o grau de severidade da catarata pode ser classificado automaticamente com técnicas de machine learning. Verificou-se que os parâmetros acústicos, velocidade e atenuação, aumentam significativamente com a progressão da catarata. Constatou-se, também, que a intensidade média do brilho das imagens de B-scan e o valor médio do parâmetro m das imagens de Nakagami, aumentam significativamente com a progressão da catarata, como resultado do aumento das regiões de alta concentração de dispersores (scatterers). As abordagens de machine learning usadas, obtiveram um bom desempenho na classificação da catarata, confirmando a existência de uma associação entre os parâmetros acústicos e o grau de severidade da catarata. No estudo realizado in vivo, foi desenvolvido um modelo animal para catarata nuclear do qual resultaram doze ratos Wistar com catarata incipiente, treze com catarata moderada e onze com catarata severa. Os sinais acústicos foram adquiridos com uma sonda oftalmológica de ultrassons de 20 MHz. A identificação dos sinais referentes às interfaces do globo ocular, requereu o desenvolvimento de um algoritmo que foi capaz de localizar o eco da cápsula posterior, de difícil detecção. Foram extraídas 27 features para detecção e classificação automática da catarata, usando os classificadores SVM, Bayes, Multilayer Perceptron e Random Forest. As durezas do núcleo e do córtex do cristalino foram medidas e a respectiva elasticidade estimada em cristalinos de 12 ratos, por técnicas de nanoindentação. Demonstrou-se que a velocidade, atenuação e o desvio de frequência (downshift) aumentam significativamente com a progressão da catarata (p < 0,001). O classificador SVM apresentou o melhor desempenho na classificação automática do grau de severidade da catarata, com uma precisão, sensibilidade e especificidade de 0,997, e um erro absoluto relativo de 0,4%. Foi observada uma diferença estatisticamente significativa para os valores da dureza entre os diferentes graus de severidade da catarata (p = 0,016). O núcleo apresentou o maior aumento da dureza com a progressão da catarata (p = 0,049). Foi obtida uma correlação moderada a boa entre a dureza do núcleo e 23 das 27 features analisadas. Foi encontrada uma forte correlação entre o módulo de Young e a dureza do cristalino (r = 0,953, para o núcleo e r = 0,701 para o córtex, p <0,001). A metodologia desenvolvida permitiu detectar a catarata nuclear in vivo em estadios iniciais, classificar automaticamente o seu grau de severidade e estimar a sua dureza de modo não invasivo. Com base nos estudos realizados, foi desenvolvido um protótipo (ESUS - Eye Scan Ultrasound System), para a deteção precoce e classificação automática da catarata, bem como para a estimação da sua dureza em tempo real. Este protótipo está protegido por um Pedido Provisório de Patente PPP108836 (2015). Cataract is a condition associated with the loss of the normal transparency of the crystalline lens. Its progression may result in partial or total vision loss. The cataract formation process is a result of fibrosis and protein aggregation within the lens. Fiber compaction and proteins aggregation have the effect of increasing the light scattering, lens opacity and hardness, leading to visual acuity loss. Currently, cataract has no pharmacological treatment or prevention and the only effective therapy to restore vision is the cataract surgical removal, replacing the natural lens by an artificial intraocular lens. The phacoemulsification is the most widely used surgical technique for cataract removal. For diagnosis and therapeutic purposes it is of major importance to identify the cataract in its early stage, and accurately determine its type and severity. The early cataract detection, the correct classification, and its hardness estimation are key factors for clinical cataract management. In particular, the cataract hardness estimation plays an important role in the optimal phacoemulsification energy level selection, which results in less surgical complications and higher safety levels. This work aims to in vivo characterize the cataract based on ultrasound techniques, namely, to early detect cataract formation, estimate its hardness, and automatically classify its severity level using machine learning. A prototype, to be validated in the future in clinical studies, was developed and tested in this work in small animals (Wistar rats). Before conducting in vivo experiments in an animal model for nuclear cataract, a proof-of-concept study was conducted ex vivo in porcine lenses. The proof-of-concept study was conducted to validate the experimental setup and to demonstrate the feasibility of the ultrasound technique for the cataract characterization noninvasively (without lens destruction). In that study the cataract was induced by lens immersion in a ethanol:2-propanol:formalin solution. A 25 MHz ultrasound transducer with a 25 mm focus and 5 mm active diameter was used to collect the acoustical parameters (velocity and attenuation), and backscattering signals. In order to visualize and characterize cataract B-scan and parametric Nakagami images were constructed. Ninety-seven features were extracted and subjected to a Principal Component Analysis. Bayes, K-Nearest-Neighbours, Fisher Linear Discriminant and Support Vector Machine (SVM) classifiers were used to automatically classify the cataract severity. The spatial distribution of protein aggregates and the fiber compaction for the different cataract degrees was also analysed based on histological slices’ images obtained in the nucleus and cortex regions. This study showed that the cataract hardness can be noninvasively estimated and the cataract severity automatically classified with machine learning techniques, using the acoustical parameters, B-scan and Nakagami parametric images, and the backscattering signals. It was observed that the acoustical parameters (velocity and attenuation) increased significantly with the cataract progression. Also, the mean brightness of the B-scan images and the mean m parameter value of the Nakagami images showed an increase over the cataract progression, as a result of the increased region of high scatterers concentration. Finally, the used machine learning techniques presented a good performance for cataract classification confirming the existence of an association between the ultrasound parameters and cataract severity. For the in vivo study, fifty rats were used: fourteen as control and thirty-six as study group. First, an animal model for nuclear cataract was developed, which provided 12 rats with incipient, 13 with moderate, and 11 with severe cataract. Then, a 20 MHz ophthalmic ultrasound probe with a focal length of 8.9 mm, and active diameter of 3 mm was used for signals collection. An algorithm was developed to identify all pulses associated to the eye boundaries, in particular the posterior capsule echo, which in general has small amplitude. Next, twenty-seven features in time and frequency domain were extracted for cataract detection and automatic classification by using SVM, Bayes, Multilayer Perceptron and Random Forest classifiers. The hardness of the nucleus and the cortex regions of the lens were also objectively measured in 12 rats using the NanoTestTM, and the respective elasticity estimated. It was shown that the velocity, attenuation and frequency downshift significantly increased with cataract progression (P<0.001). The SVM classifier showed the higher performance for the automatic classification of cataract severity, with a precision, sensitivity and specificity of 0.997, with a relative absolute error of 0.4%. A statistically significant difference was found for the hardness of the different cataract degrees (p = 0.016). The nucleus showed a higher hardness increase with cataract progression (p = 0.049). A moderate to good correlation between the features and the nucleus hardness was found in 23 out of the 27 features. A strong correlation was found between the Young’s modulus and the lens hardness (r = 0.953, for the nucleus and r = 0.701 for the cortex, p < 0.001). The methodology developed in this study allows detecting the nuclear cataract in vivo in early stages, classifying automatically its severity degree and estimating its hardness noninvasively. Following the obtained results a prototype called Eye Scan Ultrasound System (ESUS) was developed for early and automatic cataract detection and classification in real time. This prototype is currently protected under provisional patent application PPP108836 (Santos et al. 2015).

Cataract is a condition associated with the loss of the normal transparency of the crystalline lens. Its progression may result in partial or total vision loss. The cataract formation process is a result of fibrosis and protein aggregation within the lens. Fiber compaction and proteins aggregation have the effect of increasing the light scattering, lens opacity and hardness, leading to visual acuity loss. Currently, cataract has no pharmacological treatment or prevention and the only effective therapy to restore vision is the cataract surgical removal, replacing the natural lens by an artificial intraocular lens. The phacoemulsification is the most widely used surgical technique for cataract removal. For diagnosis and therapeutic purposes it is of major importance to identify the cataract in its early stage, and accurately determine its type and severity. The early cataract detection, the correct classification, and its hardness estimation are key factors for clinical cataract management. In particular, the cataract hardness estimation plays an important role in the optimal phacoemulsification energy level selection, which results in less surgical complications and higher safety levels. This work aims to in vivo characterize the cataract based on ultrasound techniques, namely, to early detect cataract formation, estimate its hardness, and automatically classify its severity level using machine learning. A prototype, to be validated in the future in clinical studies, was developed and tested in this work in small animals (Wistar rats). Before conducting in vivo experiments in an animal model for nuclear cataract, a proof-of-concept study was conducted ex vivo in porcine lenses. The proof-of-concept study was conducted to validate the experimental setup and to demonstrate the feasibility of the ultrasound technique for the cataract characterization noninvasively (without lens destruction). In that study the cataract was induced by lens immersion in a ethanol:2-propanol:formalin solution. A 25 MHz ultrasound transducer with a 25 mm focus and 5 mm active diameter was used to collect the acoustical parameters (velocity and attenuation), and backscattering signals. In order to visualize and characterize cataract B-scan and Nakagami m parameter images were constructed. Ninety-seven features were extracted and subjected to a Principal Component Analysis. Naive Bayes, K-Nearest-Neighbours, Fisher Linear Discriminant and Support Vector Machine (SVM) classifiers were used to automatically classify the cataract severity. The spatial distribution of protein aggregates and the fiber compaction for the different cataract degrees was also analysed based on histological slices’ images obtained in the nucleus and cortex regions. This study showed that the cataract hardness can be noninvasively estimated and the cataract severity automatically classified with machine learning techniques, using the acoustical parameters, B-scan and Nakagami parametric images, and the backscattering signals. It was observed that the acoustical parameters (velocity and attenuation) increased significantly with the cataract progression. Also, the mean brightness of the B-scan images and the mean value of the Nakagami m parameter showed an increase over the cataract progression, as a result of the increased region of high scatterers concentration. Finally, the used machine learning techniques presented a good performance for cataract classification (F-measure ! 0.725) confirming the existence of an association between the ultrasound parameters and cataract severity. For the in vivo study, fifty rats were used: fourteen as control and thirty-six as study group. First, an animal model for nuclear cataract was developed, which provided 12 rats with incipient, 13 with moderate, and 11 with severe cataract. Then, a 20 MHz ophthalmic ultrasound probe with a focal length of 8.9 mm, and active diameter of 3 mm was used for signals collection. An algorithm was developed to identify all pulses associated to the eye boundaries, in particular the posterior capsule echo, which in general has small amplitude. Next, twenty-seven features in time and frequency domain were extracted for cataract detection and automatic classification by using SVM, Naive Bayes, Multilayer Perceptron and Random Forest classifiers. The hardness of the nucleus and the cortex regions of the lens were also objectively measured in 12 rats using the NanoTestTM, and the respective elasticity estimated. It was shown that the velocity, attenuation and frequency downshift significantly increased with cataract progression (p < 0.001). The SVM classifier showed the higher performance for the automatic classification of cataract severity, with a precision, sensitivity and specificity of 0.997, with a relative absolute error of 0.4%. A statistically significant difference was found for the hardness of the different cataract degrees (p = 0.016). The nucleus showed a higher hardness increase with cataract progression (p = 0.049). A moderate to good correlation between the features and the nucleus hardness was found in 23 out of the 27 features. A strong correlation was found between the Young’s modulus and the lens hardness (r = 0.953, for the nucleus and r = 0.701 for the cortex, p < 0.001). The methodology developed in this study allows detecting the nuclear cataract in vivo in early stages, classifying automatically its severity degree and estimating its hardness noninvasively. Following the obtained results a prototype called Eye Scan Ultrasound System (ESUS) was developed for early and automatic cataract detection and classification in real time. This prototype is currently protected under provisional patent application PPP108836 (Santos et al. 2015).