Worldwide, 537 million people aged between 20 and 79 were living with diabetes in 2021 and this number is predicted to rise to 643 million by 2030 and 783 million by 2045. Diabetes is responsible for 6.7 million deaths in 2021 - 1 every 5 seconds. In order to contain the growth of these numbers and improve the life quality of patients, it is important to screen categories at risk, diagnosing diabetes at its early stage. To do so, one of the most popular tests is the glycated hemoglobin (HbA1C) test, that measures the percentage of blood sugar attached to hemoglobin and shows the average blood sugar level for the past 2 to 3 months. This test requires a blood sample acquisition, which is an invasive procedure. For this reason, predicting HbA1c from a fundus camera image can be useful, and it also provides other retinal biomarkers. In this work, it has been tested whether this can be done with enough accuracy using convolutional neural networks. To do so, a dataset made of 100153 images coming from 16799 subjects has been split in training, validation and test set to train and test an EfficientNet-B2 convolutional neural network. For each subject, many different images were taken over time for both eyes together with an HbA1c measurement. Two different experiments have been performed: the first one trying to predict the latest HbA1c measurement using the first image collected for each subject and the second one trying to predict the HbA1c value corresponding to each image. After having found that the first experiment was unfeasible, due to the misconception of predicting future information from a tissue carrying information about the past history, the focus shifted to the second experiment. In this case, even if the results were not excellent, they have shown more promising perspectives and coherence of the predictions over time. An analysis of the effect of sex and age on the cumulative HbA1c value as been performed, confirming that these variables do not affect it. Afterwards, an analysis of the time trend of the predictions has been performed, fitting them with a linear model and extracting its parameters, for each subject. Depending on the position of the predicted fit with the respect to the actual fit, the subjects have been given three categories. A $\chi^2$ test has been performed to inspect whether there was an association or not between these categories and the death outcome. The same has been done for a risk-class category, build on whether the predicted slope was greater or smaller than the actual one. Since this work is strongly experimental and one of the first of his kind, it has the aim to pave the path in this specific field and has big room for improvement. Most of the limitations of this work come from the cumulative nature of the data, but some suggestions on how to improve already exist and are presented in the dedicated chapter.
Nel 2021, in tutto il mondo, 537 milioni di persone di età compresa tra i 20 e i 79 anni erano malate di diabete. Tale numero è previsto crescere fino a 643 milioni nel 2030 e 783 milioni nel 2045. Nel 2021 il diabete è stato responsabile della morte di 6.7 milioni di persone. Per contenere la crescita di questi numeri e migliorare la qualità della vita dei pazienti, è importante rilevare le categorie a rischio e diagnosticare il diabete alle sue fasi iniziali. Per farlo, uno dei test più popolari è quello dell’emoglobina glicata (HbA1c), che misura la percentuale di glucosio nel sangue legato all’emoglobina e indica il livello medio di glucosio del sangue negli ultimi 2/3 mesi. Questo test richiede il prelievo e l’analisi di un campione di sangue, procedura invasiva per il paziente. Per questo motivo, predire l’emoglobina glicata da un’immagine di fundus camera può essere utile, oltre che per fornire altri biomarcatori. In questo lavoro è stato verificato se tale task fosse possibile e con quali performance, usando una rete neurale convoluzionale. Per farlo è stato usato un dataset di 100153 immagini provenienti da 16799 soggetti ed EfficientNet-B2 come rete neurale. Per ciascun soggetto sono disponibili più immagini di entrambi gli occhi raccolte nel tempo, con la rispettiva misura di HbA1c cumulativa. Sono stati svolti due diversi esperimenti: il primo tentando di predire l’ultimo valore di HbA1c cumulativa dalla prima immagine e il secondo tentando di predire il valore corrispondente a ciascuna immagine. Dopo aver realizzato che il primo esperimento non era fattibile, in quanto la retina è un tessuto che presenta features di eventi passati e non futuri, gli sforzi si sono concentrati sul secondo esperimento. Anche se i risultati non sono eccellenti, questo esperimento ha dimostrato una forte coerenza del trend delle predizioni nel tempo. Un’analisi dell’effetto del sesso e dell’età è stata svolta, per confermare che queste variabili non hanno associazione con l’emoglobina glicata cumulativa. In seguito è stata svolta un’analisi del trend delle predizioni nel tempo, fittandole con un modello lineare ed estraendone i parametri. Questi sono stati usati per creare due variabili categoriche, la cui associazione con l’outcome di morte del paziente è stata testata con un test statistico del chi quadro. Dato che questo lavoro è sperimentale e uno dei primi sull’emoglobina glicata cumulativa, e ha lo scopo di aprire la via in questo settore specifico e ha molta prospettiva di miglioramento. Molte delle limitazioni di questo studio sono collegate proprio alla cumulatività delle misure, ma alcuni suggerimenti su come superarli sono presentati nel capitolo dedicato.
Stima dell'esposizione all'emoglobina glicata da immagini retiniche usando deep learning in una popolazoine diabetica
QUINTO, ANDREA
2021/2022
Abstract
Worldwide, 537 million people aged between 20 and 79 were living with diabetes in 2021 and this number is predicted to rise to 643 million by 2030 and 783 million by 2045. Diabetes is responsible for 6.7 million deaths in 2021 - 1 every 5 seconds. In order to contain the growth of these numbers and improve the life quality of patients, it is important to screen categories at risk, diagnosing diabetes at its early stage. To do so, one of the most popular tests is the glycated hemoglobin (HbA1C) test, that measures the percentage of blood sugar attached to hemoglobin and shows the average blood sugar level for the past 2 to 3 months. This test requires a blood sample acquisition, which is an invasive procedure. For this reason, predicting HbA1c from a fundus camera image can be useful, and it also provides other retinal biomarkers. In this work, it has been tested whether this can be done with enough accuracy using convolutional neural networks. To do so, a dataset made of 100153 images coming from 16799 subjects has been split in training, validation and test set to train and test an EfficientNet-B2 convolutional neural network. For each subject, many different images were taken over time for both eyes together with an HbA1c measurement. Two different experiments have been performed: the first one trying to predict the latest HbA1c measurement using the first image collected for each subject and the second one trying to predict the HbA1c value corresponding to each image. After having found that the first experiment was unfeasible, due to the misconception of predicting future information from a tissue carrying information about the past history, the focus shifted to the second experiment. In this case, even if the results were not excellent, they have shown more promising perspectives and coherence of the predictions over time. An analysis of the effect of sex and age on the cumulative HbA1c value as been performed, confirming that these variables do not affect it. Afterwards, an analysis of the time trend of the predictions has been performed, fitting them with a linear model and extracting its parameters, for each subject. Depending on the position of the predicted fit with the respect to the actual fit, the subjects have been given three categories. A $\chi^2$ test has been performed to inspect whether there was an association or not between these categories and the death outcome. The same has been done for a risk-class category, build on whether the predicted slope was greater or smaller than the actual one. Since this work is strongly experimental and one of the first of his kind, it has the aim to pave the path in this specific field and has big room for improvement. Most of the limitations of this work come from the cumulative nature of the data, but some suggestions on how to improve already exist and are presented in the dedicated chapter.File | Dimensione | Formato | |
---|---|---|---|
Quinto_Andrea.pdf
accesso aperto
Dimensione
12.16 MB
Formato
Adobe PDF
|
12.16 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/42439