Impact of model selection procedure on Deep Neural Networks ensemble for the Choroid Plexus segmentation in Multiple Sclerosis

The Choroid Plexus (ChP) is a brain vascular tissue involved in regulatory processes. ChP Volume (ChPV) modifications are related to neurodegenerative disorders. Therefore, ChPV, that can be obtained from manual segmentation of brain MRI, is an imaging biomarker candidate to monitor disease evolution. This work proposes a method for the automatic segmentation of ChP based on hyperparameters optimization of Deep Neural Networks (DNNs). Twenty-Seven hyperparameters and architectures combinations were trained on T1-w MRI with two different selection strategies: select the best models using the routinely used Dice Coefficient and combining it to the Absolute Percentage Volume Difference. The selection of the ten best models was made on bias and variance of Absolute Percentage Volume Difference and best DNNs were ensembled by majority voting for both selection strategies. The proposed ensemble models outperform single DNNs (Dice Coefficient for both ensembles: 0.81±0.07; Percentage Volume Difference - ensemble Dice: 0.41±10.75%; ensemble Dice&Volume: -0.05±10.49%). Ensemble segmentations obtained using the combination of Dice and Absolute Percentage Volume Difference are preferable since the variance obtained in the testing phase is slightly lower than the commonly used Dice metric. Therefore, the proposed ensemble of DNNs, selected exploiting both Dice and Absolute Percentage Volume Difference, is a promising tool to obtain automatic quantification of the ChPV. © 2023 Convegno Nazionale di Bioingegneria. All rights reserved.