Predicting response to neo-adjuvant chemotherapy of liver metastases (mts) using CT images is of key importance to provide personalized treatments. However, manual segmentation of mts should be avoid to develop methods that could be integrated into the clinical practice. The aim of this study is to evaluate if and how much automatic segmentation can affect a radiomics-based method to predict response to neoadjuvant chemotherapy of individual liver mts. To this scope, we developed an automatic deep learning method to segment liver mts, based on the U-net architecture, and we compared the classification results of a classifier fed with manual and automatic masks. In the validation set composed of 39 liver mts, the automatic deeplearning algorithm was able to detect 82% of mts, with a median precision of 67%. Using manual and automatic masks, we obtained the same classification in 19/32 mts. In case of mts with largest diameter > 20 mm, the precision of the segmentation does not impact the classification results and we obtained the same classification with both masks. Conversely, with smaller mts, we showed that a Dice coefficient of at least 0.5 should be obtained to extract the same information from the two segmentations. This are very important results in the perspective of using radiomics-based approach to predict response to therapy into clinical practice. Indeed, either precisely manually segment all lesions or refine them after automatic segmentation is a time-consuming task that cannot be performed on a daily basis.
Deep learning to segment liver metastases on CT images: Impact on a radiomics method to predict response to chemotherapy
Giannini V.;Defeudis A.;Cappello G.;Vassallo L.;Mazzetti S.;Panic J.;Regge D.;
2020-01-01
Abstract
Predicting response to neo-adjuvant chemotherapy of liver metastases (mts) using CT images is of key importance to provide personalized treatments. However, manual segmentation of mts should be avoid to develop methods that could be integrated into the clinical practice. The aim of this study is to evaluate if and how much automatic segmentation can affect a radiomics-based method to predict response to neoadjuvant chemotherapy of individual liver mts. To this scope, we developed an automatic deep learning method to segment liver mts, based on the U-net architecture, and we compared the classification results of a classifier fed with manual and automatic masks. In the validation set composed of 39 liver mts, the automatic deeplearning algorithm was able to detect 82% of mts, with a median precision of 67%. Using manual and automatic masks, we obtained the same classification in 19/32 mts. In case of mts with largest diameter > 20 mm, the precision of the segmentation does not impact the classification results and we obtained the same classification with both masks. Conversely, with smaller mts, we showed that a Dice coefficient of at least 0.5 should be obtained to extract the same information from the two segmentations. This are very important results in the perspective of using radiomics-based approach to predict response to therapy into clinical practice. Indeed, either precisely manually segment all lesions or refine them after automatic segmentation is a time-consuming task that cannot be performed on a daily basis.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
