Caruso, Martina (2023) Predizione preoperatoria di metastasi dei linfonodi ascellari da carcinoma mammario utilizzando il machine learning applicato ad immagini ecografiche B-mode e di sonoelastografia shear-wave. [Tesi di dottorato]

Anteprima

Testo Caruso_Martina_36.pdf Download (1MB) | Anteprima

Abstract

Purpose 1) To build and validate an US-based machine-learning (ML) model to predict axillary lymph node (ALN) status in breast cancer (BC) in a multicentric setting. 2) To assess whether a ML algorithm could empower the ability of US with the addition of shear-wave elastography (SWE) to preoperatively define ALN status in BC in a single-institution setting. Methods and materials Patients with at least one histologically proven BC lesion, who underwent preoperative breast US, were retrospectively enrolled in three different Institutions (multicenter study). Furthermore, patients with preoperative breast US integrated with SWE were retrospectively selected in only one Institution (monocentric study). BC lesions were segmented on US and SWE images by three different operators and radiomics features (first, second, higher order) were extracted. Regarding the multicenter study, a multi-step US feature selection was performed using Intraclass Correlation Coefficient (ICC) analysis and principal component analysis (PCA). Thereafter, a Random Forest (RF) ML classifier was applied to the dataset to predict the ALN status (positive/negative for metastasis) and its performance assessed through the Matthews Correlation Coefficient (MCC). Regarding the monocentric study, a multi-step US and SWE feature selection was performed using ICC, intercorrelation analysis, and information gain analysis. A Simple Logistic ML classifier was applied to the dataset to predict the ALN status and its performance assessed through the Area Under the ROC curve (AUC) and MCC. Results A total of 293 BC lesions (ALN negative: 176; ALN positive: 117) were included in the multicenter study. Three datasets were identified as follows: 1) Training set, composed of 233 BCs (ALN-: 140; ALN+: 93); 2) validation set including 30 BCs (ALN-: 17; ALN+: 13); and 3) test set made of 30 BCs (ALN-: 18; ALN+: 12). 549 radiomics features were extracted from US images; of these, 280 were discarded according to ICC analysis, with a total of 5 features finally selected by PCA. RF classifier showed a MCC of 0.97, 0.11 and 0.08 in the training, validation, and test set, respectively. A total of 133 BC lesions (ALN-: 76; ALN+: 57) were included in the monocentric study and divided into two sets: 1) training set, composed of 89 BC lesions (ALN-: 52; ALN+: 37); and 2) test set, including 44 BC lesions (ALN-: 24; ALN+: 20). 1098 radiomics features were extracted from US and SWE images (549 features from each set); of these, 835 were discarded according to ICC analysis and 241 according to intercorrelation analysis. 8 features of the 22 remaining were selected through the information gain analysis. Simple Logistic classifier showed AUC of 0.685 and 0.677, a MCC of 0.387 and 0.375 in the training and test set, respectively. Limitations The relatively small patient population and, for the monocentric study, the lack of an external validation set from another Institution. Conclusion ML applied showed promises in the preoperative assessment of ALN status in BC, also with the addition of SWE images. However, further efforts are necessary to improve the generalizability of the model when applied to external datasets.

Downloads

Actions (login required)

Modifica documento