Regulerization of Convolutional-Neural-Networks Models for Echocardiographic Left-Ventricle Segmentation using Deep-Sparse-Coding Networks

Background: Exact measures of the left-ventricular motion are highly sought-after for clinical diagnosis. Currently, tracking heart wall-motion requires automatic segmentation of a single frame. CNNs for echocardiac segmentation have been suggested for this task. Recently, attention is given to improving regularization of Segmentation CNNs (SCNN), to reduce model ambiguity and increase accuracy and generalization. Constraining a CNN model is strictly non-trivial to formulize as penalty measures on salient priors (as a function of model parameters). Using a Convolutional-Auto-Encoder was suggested for encoding latent representations of segmentations, allowing an additional error measure between output and ground-truth, expressing distance in latent features. Adding the measure as a regularization term for the SCNN model showed improved convergence and higher accuracy on echocardiographic data. Using a Supervised-Deep-Sparse-Coding-Network (SDSCN), recently proposed as a new deep-learning network, we trained a supervised encoder and utilized it to train a uniquely regularized SCNN.

Methods: we implemented a framework on Matlab for SCNN training. The framework includes a CAE encoder and a SDSCN encoder for regularization. A dataset of ~140 echocardiac hand-labeled frames was used for training.

Results: both encoders allowed comparable improvement in SCNN convergence rates and accuracies, while the new SDSCN encoder utilizes only ~40K parameters, versus ~230K for the CAE, allowing for a much better generalization.

Conclusions: A SDSCN encoder can allow better generalization, opacity and accuracy for SCNN models in echocardiac segmentation. We plan to design an advanced segmentation framework utilizing supervised encoders. The center model is expected to capture meaningful and provable anatomical sense.