Primary Central System lymphoma (PCNSL) is a rare and aggressive form of extranodal non-Hodgkin lymphoma involving the brain, leptomeninges, eye or spinal cord. This study aims to support clinician in the management of that disease through the design of a machine learning based classification method using radiomics features of multiparametric MRI to correctly predict the Overall Survival (OS) in the short- and long-term. Multiparametric MRI of 59 patients diagnosed with PCNSL treated with high dose methotrexate- (HDMTX) based chemotherapy were used. Patients were stratified into two groups: those with long-term OS (> 12 months) and those with short-term OS (< 12 months). Each patient has T1, contrast enhanced T1, T2 and FLAIR MRI. MRI images underwent a pre-processing phase including bias correction, skull stripping, registration, Z-score normalization, and voxels resampling (1mm3). The lesion was segmented by an expert radiologist and the tumor contours were used to extract features from the multiparametric MRI. Once these features have been extracted, the most significant and informative were selected through a Recursive Feature Elimination (RFE) algorithm to train three classifiers: Logistic Regression, Support Vector Machine (SVM) and Random Forest. A 5-fold stratified repeated cross-validation were applied to the dataset to achieve the classification task. The prediction accuracy, F1-score, and area under the ROC curve (AUC) were obtained for each fold to evaluate the performance of the ma-chine learning based classification. Logistic Regression and SVM classifiers reach higher performances with respect Random Forest classifiers (accuracy of 0.76 ± 0.1, F1-score of 0.80 ± 0.1 and AUC of almost 0.80 ± 0.1 for the first, 0.72 ± 0.1, 0.76 ± 0.1 and 0.80 ± 0.1 for the latter). The Z-score normalization has been demonstrated to increase classification performances on T1 extracted features of the 10% for all the classifiers. Radiomic features were compared between the two groups considering the effects of gender and age. Overall, 19 and 2 radiomics features were recognized respectively on T1 and T2 sequence as discriminating between the classes while no discriminating features have been found for the FLAIR sequence.
Il linfoma primario del sistema centrale (PCNSL) è una forma rara e aggressiva di linfoma non-Hodgkin extranodale che coinvolge cervello, leptomeningi, occhio o midollo spinale. Questo studio mira a sviluppare un metodo di classificazione basato sull’utilizzo di feature radiomiche estratte da MRI multiparametrica per la predizione della sopravvivenza globale (OS) a breve e lungo termine. Sono stati reclutati 59 pazienti con diagnosi di PCNSL trattati con chemioterapia basata sull’uso di metrotrexato ad alte dosi (HDMTX). I pazienti sono stati stratificati in due gruppi: quelli con OS a lungo termine (> 12 mesi) e quelli con OS a breve termine (< 12 mesi). Ogni paziente è stato sottoposto a risonanza magnetica T1, T1 con contrasto, T2 e FLAIR in tre fasi del trattamento. Le immagini MRI sono state sottoposte a una fase di pre-processing che comprende la correzione delle disomogeneità del campo magnetico, lo stripping del cranio, la registrazione, la normalizzazione con Z-score e il ricampionamento dei voxel (1mm3). La lesione è stata segmentata da un radiologo esperto e i contorni del tumore sono stati utilizzati per estrarre le feature radiomiche dalle MRI pretrattamento. Una volta estratte le feature radiomiche, quelle più significative sono state selezionate attraverso un algoritmo Recursive Feature Elimination (RFE) per addestrare tre classificatori: Regressione logistica, Support Vector Machine (SVM) e Random Forest. Una crossvalidazione ripetuta e stratificata su 5 fold è stata utilizzata in fase di addestramento. L'accuratezza della classificazione, il punteggio F1 e l'area sotto la curva ROC (AUC) sono stati calcolati in ogni fold per valutare le prestazioni della classificazione. I classificatori Regressione logistica e SVM hanno raggiunto prestazioni più elevate rispetto al classificatore Random Forest (precisione di 0,76 ± 0,1, F1-score di 0,80 ± 0,1 e AUC di quasi 0,80 ± 0,1 per il primo, 0,72 ± 0,1, 0,76 ± 0,1 0,80 ± 0,1). La normalizzazione dello Z-score ha dimostrato aumentare le prestazioni di classificazione sulle feature estratte dalla sequenza T1 del 10% per tutti i classificatori. Le feature radiomiche sono state con-frontate tra i due gruppi considerando gli effetti del sesso e dell'età. Nel complesso, 18 e 2 feature sono state riconosciute rispettivamente sulla sequenza T1 e T2 come discriminanti tra le classi, mentre nessuna feature discriminante è stata identificata nella sequenza FLAIR.
Overall survival prediction in PCNSL with radiomic features and machine learning
LODDO, MONICA
2021/2022
Abstract
Primary Central System lymphoma (PCNSL) is a rare and aggressive form of extranodal non-Hodgkin lymphoma involving the brain, leptomeninges, eye or spinal cord. This study aims to support clinician in the management of that disease through the design of a machine learning based classification method using radiomics features of multiparametric MRI to correctly predict the Overall Survival (OS) in the short- and long-term. Multiparametric MRI of 59 patients diagnosed with PCNSL treated with high dose methotrexate- (HDMTX) based chemotherapy were used. Patients were stratified into two groups: those with long-term OS (> 12 months) and those with short-term OS (< 12 months). Each patient has T1, contrast enhanced T1, T2 and FLAIR MRI. MRI images underwent a pre-processing phase including bias correction, skull stripping, registration, Z-score normalization, and voxels resampling (1mm3). The lesion was segmented by an expert radiologist and the tumor contours were used to extract features from the multiparametric MRI. Once these features have been extracted, the most significant and informative were selected through a Recursive Feature Elimination (RFE) algorithm to train three classifiers: Logistic Regression, Support Vector Machine (SVM) and Random Forest. A 5-fold stratified repeated cross-validation were applied to the dataset to achieve the classification task. The prediction accuracy, F1-score, and area under the ROC curve (AUC) were obtained for each fold to evaluate the performance of the ma-chine learning based classification. Logistic Regression and SVM classifiers reach higher performances with respect Random Forest classifiers (accuracy of 0.76 ± 0.1, F1-score of 0.80 ± 0.1 and AUC of almost 0.80 ± 0.1 for the first, 0.72 ± 0.1, 0.76 ± 0.1 and 0.80 ± 0.1 for the latter). The Z-score normalization has been demonstrated to increase classification performances on T1 extracted features of the 10% for all the classifiers. Radiomic features were compared between the two groups considering the effects of gender and age. Overall, 19 and 2 radiomics features were recognized respectively on T1 and T2 sequence as discriminating between the classes while no discriminating features have been found for the FLAIR sequence.File | Dimensione | Formato | |
---|---|---|---|
Loddo_Monica.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
21.88 MB
Formato
Adobe PDF
|
21.88 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/10589/183118