Unstable carotid atherosclerosis is assessed by World Health Organization as major cause of stroke, which is leading cause of death and disability in the United States [1]. Current clinical treatment of carotid atherosclerosis fails to prevent many strokes: only 1 is prevented over 20 patients treated by surgery [2]; this may be related to the fact that patient risk assessment is performed only considering the percentage of vessel stenosis. The goal of this work is to improve the accuracy of risk assessment and overall treatment of carotid atherosclerosis disease to prevent stroke by developing an automated classification tool to support decision-making, which analyzes the histological composition of the carotid artery to study all the cellular conditions that might differ between symptomatic and asymptomatic patients. On this purpose, the availability of histological images of carotid sections for 128 patients with H&E and Trichrome staining at 40X magnification let us obtain a high detailed description of the entire plaque cellular components. Hence, we hypothesized that Deep Learning models, like pretrained convolutional neural networks, would have allowed to identify, over this high resolution image dataset, the key characteristics to be then studied by Machine Learning methods to discriminate patients at high and low risk of stroke. According to it, we zoomed and focused over low-level cellular shapes by building an extended grid of sub images over each carotid section; from each of them we extracted a set of features and made over them a hierarchical clustering in order to create a count matrix describing the composition of each section; we then used it as input to a classifier built with the mixed effects Random Forests model and managed to be able to differentiate patients at high and low risk of stroke. For both staining we reached significant results, with accuracy in prediction around 98%. At last, we’ve carried on a last analysis to rank and thus highlight the most discriminative histological dynamics, useful to better comprehend the cellular processes involved in atherosclerosis. The set of all of the analysis techniques developed this project could have a strong impact in clinical practice because they might potentially lead to a big improvement of the therapeutic strategies used for prevention, mortality due to unneeded surgeries, prevention of associated morbidity and cost/benefit ratios. [1] Feigin VL, Forouzanfar MH, Krishnamurthi R, et al. Global and regional burden of stroke during 1990-2010: findings from the Global Burden of Disease Study 2010 [published correction appears in Lancet. 2014 Jan 18;383(9913):218]. Lancet. [2] Halliday A, Harrison M, Hayter E, et al. 10-year stroke prevention after successful carotid endarterectomy for asymptomatic stenosis (ACST-1): a multicentre randomised trial. Lancet. 2010;376(9746):1074–1084.
CLASSIFICAZIONE AUTOMATICA DEI PAZIENTI AD ALTO RISCHIO DI ICTUS USANDO IMMAGINI ISTOLOGICHE DELLE CAROTIDI. L'aterosclerosi carotidea instabile è valutata dall'Organizzazione mondiale della sanità come principale causa di ictus; essa è responsabile per un elevato numero di cause di morte e disabilità negli Stati Uniti [1]. L'attuale trattamento clinico dell'aterosclerosi carotidea non riesce a prevenire molti ictus: solo 1 è prevenuto su 20 pazienti trattati con chirurgia [2]; ciò può essere correlato al fatto che la valutazione del rischio del paziente viene eseguita solo considerando la percentuale di stenosi del vaso. L'obiettivo di questo lavoro è quello di migliorare l'accuratezza della valutazione del rischio e il trattamento complessivo della malattia aterosclerotica carotidea per prevenire l'ictus sviluppando uno strumento di classificazione automatizzata a supporto del processo decisionale, che analizza la composizione istologica dell'arteria carotidea per studiare tutte le condizioni cellulari che potrebbero differire tra pazienti sintomatici e asintomatici. A tal fine, la disponibilità di immagini istologiche di sezioni di carotidi per 128 pazienti con colorazione H&E e Tricrome con ingrandimento 40X ci ha permesso di ottenere una descrizione dettagliata di tutti i componenti cellulari della placca. Da qui abbiamo ipotizzato che i modelli di Deep Learning, come le reti neurali convoluzionali pre-allenate, avrebbero permesso di identificare, su questo dataset di immagini ad alta risoluzione, le caratteristiche chiave da studiare poi con i metodi di Machine Learning per discriminare i pazienti ad alto e basso rischio di ictus. A tal proposito, abbiamo compiuto operazioni di focus sulle forme cellulari di basso livello costruendo un’estesa griglia di sotto-immagini su ogni sezione carotidea; da ciascuna di esse abbiamo estratto un insieme di features e poi eseguito un clustering gerarchico al fine di creare una matrice di conteggio che potesse descrivere la composizione di ciascuna sezione; essa è stata quindi usata come input per un classificatore costruito con il modello mixed effects Random Forests, che ci ha permesso di differenziare i pazienti ad alto e basso rischio di ictus. Per entrambe le colorazioni abbiamo raggiunto risultati significativi, con un'accuratezza in predizione intorno al 98%. Infine, abbiamo effettuato un'ultima analisi per stilare un ranking delle dinamiche istologiche con più alto potere discriminatorio, utile per evidenziare e comprendere meglio i processi cellulari coinvolti nell'aterosclerosi. L'insieme di tutte le tecniche di analisi sviluppate in questo progetto potrebbero avere un forte impatto nella pratica clinica in quanto porterebbero potenzialmente a un grande miglioramento delle strategie terapeutiche utilizzate per la prevenzione, della mortalità a causa di interventi chirurgici non necessari, della prevenzione delle patologie associate e dei rapporti costo/beneficio. [1] Feigin VL, Forouzanfar MH, Krishnamurthi R, et al. Global and regional burden of stroke during 1990-2010: findings from the Global Burden of Disease Study 2010 [published correction appears in Lancet. 2014 Jan 18;383(9913):218]. Lancet. [2] Halliday A, Harrison M, Hayter E, et al. 10-year stroke prevention after successful carotid endarterectomy for asymptomatic stenosis (ACST-1): a multicentre randomised trial. Lancet. 2010;376(9746):1074–1084.
AUTOMATIC CLASSIFICATION OF PATIENTS AT HIGH RISK OF STROKE USING CAROTIDS HISTOLOGICAL IMAGES
BARIGGI, EDOARDO
2018/2019
Abstract
Unstable carotid atherosclerosis is assessed by World Health Organization as major cause of stroke, which is leading cause of death and disability in the United States [1]. Current clinical treatment of carotid atherosclerosis fails to prevent many strokes: only 1 is prevented over 20 patients treated by surgery [2]; this may be related to the fact that patient risk assessment is performed only considering the percentage of vessel stenosis. The goal of this work is to improve the accuracy of risk assessment and overall treatment of carotid atherosclerosis disease to prevent stroke by developing an automated classification tool to support decision-making, which analyzes the histological composition of the carotid artery to study all the cellular conditions that might differ between symptomatic and asymptomatic patients. On this purpose, the availability of histological images of carotid sections for 128 patients with H&E and Trichrome staining at 40X magnification let us obtain a high detailed description of the entire plaque cellular components. Hence, we hypothesized that Deep Learning models, like pretrained convolutional neural networks, would have allowed to identify, over this high resolution image dataset, the key characteristics to be then studied by Machine Learning methods to discriminate patients at high and low risk of stroke. According to it, we zoomed and focused over low-level cellular shapes by building an extended grid of sub images over each carotid section; from each of them we extracted a set of features and made over them a hierarchical clustering in order to create a count matrix describing the composition of each section; we then used it as input to a classifier built with the mixed effects Random Forests model and managed to be able to differentiate patients at high and low risk of stroke. For both staining we reached significant results, with accuracy in prediction around 98%. At last, we’ve carried on a last analysis to rank and thus highlight the most discriminative histological dynamics, useful to better comprehend the cellular processes involved in atherosclerosis. The set of all of the analysis techniques developed this project could have a strong impact in clinical practice because they might potentially lead to a big improvement of the therapeutic strategies used for prevention, mortality due to unneeded surgeries, prevention of associated morbidity and cost/benefit ratios. [1] Feigin VL, Forouzanfar MH, Krishnamurthi R, et al. Global and regional burden of stroke during 1990-2010: findings from the Global Burden of Disease Study 2010 [published correction appears in Lancet. 2014 Jan 18;383(9913):218]. Lancet. [2] Halliday A, Harrison M, Hayter E, et al. 10-year stroke prevention after successful carotid endarterectomy for asymptomatic stenosis (ACST-1): a multicentre randomised trial. Lancet. 2010;376(9746):1074–1084.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/19762