Interest in applying machine learning techniques for medical image classification and segmentation continues to grow at a rapid pace and – as a consequence – models are starting to be developed and deployed for clinical applications. The field of machine learning has the potential to fundamentally improve healthcare systems by capitalizing on the advances of deep learning in computer vision, natural language processing, and speech recognition, facilitated by the increasing accumulation of medical data and the widening ability of computing resources. In this work we build a 3D deep learning convolutional neural network (CNN) and we use it for two main scopes: first we analyze how the manufacturer distribution can represent a source of bias for CNN models and so that this distribution should be taken into account when building datasets for retrospective studies; second we use this network to solve a real clinical problem, the classification of idiopathic Normal Pressure Hydrocephalous (iNPH). For the first experiments on the manufacturer distribution we first demonstrate that a CNN can learn to distinguish the imaging scanner manufacturer and then that this bias can substantially impact model performance, as well as it is proportional to the difficulty of the task. The recognition of such bias is critical to develop robust, generalizable models that will be crucial for clinical application in world-real data distribution. In the second part we show that the CNN could assist the difficult radiological diagnosis of iNPH on head CT (with an AUROC of 0.96 on the test-set), lessening misdiagnosis which can have a very important clinical consequence as shunting in these patients often relieves all symptoms. The gold standard label was based upon the electronic health record metadata meeting the Japanese Society for the Normal Pressure Hydrocephalus 2nd edition diagnostic criteria for definite iNPH. This work highlights two main aspects of deep learning models when used to solve clinical problems: first it shows that data need to carefully selected and that particular attention should be payed to detect, quantify and possibly attenuate all sources of bias; second, it shows how these new tools can help clinicians by indicating a probable diagnosis to them, leading to patient follow-up and consequently to better care.
L’interesse in applicare algoritmi di machine learning per la classificazione e la segmentazione di immagini biomediche continua a crescere e – come conseguenza – questi modelli stanno iniziando ad essere sviluppati e utilizzati per applicazioni cliniche. Il campo del machine learning ha il potenziale di migliorare drasticamente I sistemi sanitari capitalizzando sui progressi del deep learning nel computer vision, natural language processing e speech recognition, facilitati dal continuo accumularsi di dati biomedicali e l’ampiamento delle abilità delle risorse computazionali. In questo lavoro di tesi costruiamo un 3D deep learning convolutional neural network e lo usiamo per due scopi principali: prima analizziamo come la distribuzione di fabbricanti di macchinari può rappresentare una sorgente di bias per CNN, quindi sottolineiamo quanto importante sia tenere conto di questa distribuzione quando si collezionano dati in un dataset per studi retrospettivi; nella seconda parte usiamo questo network per risolvere un problema reale clinico: la classificazione del idrocefalo idiopatico normoteso (iNPH). Per i primi esperimenti sulla distribuzione di fabbricanti di macchinari, prima dimostriamo che una CNN può facilmente imparare a distinguere le immagini provenienti da uno scanner di un fabbricante o di un altro, poi vediamo come questo bias può impattare sostanzialmente la performance del modello come anche è proporzionale alla difficoltà della classificazione. Il riconoscimento di questo bias è critico nello sviluppo di modelli robusti e generalizzabili ed è cruciale per applicazioni cliniche con distribuzioni reali e non ideali. Nella seconda parte presenteremo che la CNN sviluppata può assistere la difficile diagnosi radiologica del iNPH su TAC della testa (con un AUROC di 0.96 nel test-set), diminuendo la misdiagnosi che può avere una conseguenza clinica molto importante quale applicare uno shunt che in questi pazienti spesso migliora tutti i sintomi. Il gold standard è stato deciso con I record elettronici (EHR) che soddisfano I criteri diagnostici per iNPH definitive della Japanese Society for the Normal Pressure Hydrocephalus 2nd edition. Questo lavoro vuole evidenziare due principali aspetti dei modelli di deep learning quando usati per risolvere problemi clinici: prima di tutto che i dati devono essere selezionati con cautela e che bisogna prestare particolare attenzione nel rilevare, quantificare e possibilmente attenuare tutte le fonti di bias; come seconda cosa mostra come questi nuovi strumenti possono aiutare i clinici indicando loro una probabile diagnosi, portando a follow-up del paziente e di conseguenza ad una cura migliore.
3D CNN for detection of idiopathic normal pressure hydrocephalus and for evaluation of manufacturer bias using non contrast head CT
Biondetti, Giorgio Pietro
2019/2020
Abstract
Interest in applying machine learning techniques for medical image classification and segmentation continues to grow at a rapid pace and – as a consequence – models are starting to be developed and deployed for clinical applications. The field of machine learning has the potential to fundamentally improve healthcare systems by capitalizing on the advances of deep learning in computer vision, natural language processing, and speech recognition, facilitated by the increasing accumulation of medical data and the widening ability of computing resources. In this work we build a 3D deep learning convolutional neural network (CNN) and we use it for two main scopes: first we analyze how the manufacturer distribution can represent a source of bias for CNN models and so that this distribution should be taken into account when building datasets for retrospective studies; second we use this network to solve a real clinical problem, the classification of idiopathic Normal Pressure Hydrocephalous (iNPH). For the first experiments on the manufacturer distribution we first demonstrate that a CNN can learn to distinguish the imaging scanner manufacturer and then that this bias can substantially impact model performance, as well as it is proportional to the difficulty of the task. The recognition of such bias is critical to develop robust, generalizable models that will be crucial for clinical application in world-real data distribution. In the second part we show that the CNN could assist the difficult radiological diagnosis of iNPH on head CT (with an AUROC of 0.96 on the test-set), lessening misdiagnosis which can have a very important clinical consequence as shunting in these patients often relieves all symptoms. The gold standard label was based upon the electronic health record metadata meeting the Japanese Society for the Normal Pressure Hydrocephalus 2nd edition diagnostic criteria for definite iNPH. This work highlights two main aspects of deep learning models when used to solve clinical problems: first it shows that data need to carefully selected and that particular attention should be payed to detect, quantify and possibly attenuate all sources of bias; second, it shows how these new tools can help clinicians by indicating a probable diagnosis to them, leading to patient follow-up and consequently to better care.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_final.pdf
solo utenti autorizzati dal 15/09/2023
Dimensione
8.77 MB
Formato
Adobe PDF
|
8.77 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/166788