Sammendrag
It has been shown that deep neural networks can accuratly segment the left ventricle (LV), myocardium and left atrium in apical two and four chamber (A2C and A4C) views. While segmentation of apical long-axis (ALAX) views is quite similar to A2C and A4C, there is one major difference; the left ventricular outflow tract (LVOT) which restricts the myocardium. The objectives of this work were to accurately segment ALAX views, investigate if transfer learning from A2C/A4C improves accuracy, and study how a single network can learn to segment all three views.The CAMUS dataset of 500 patients together with an additional dataset of 106 patients with ALAX views were used for training and testing using 10-fold cross-validation. The results showed that by training from scratch the neural network was able to segment ALAX views, but with a lower accuracy to that of A2C/A4C views. Transfer learning only slightly improved mycoardium accuracy (0.77 to 0.78), but was statistically significant (p-value 0.001). Multi-view segmentation with the baseline network showed a reduction in accuracy, resulting in 38 cases of incorrect segmentations in terms of LVOT. The proposed network reduced the number of incorrect segmentations to 8, and achieved the best overall accuracy in terms of dice score where the improvement in myocardium segmentation accuracy (0.776 to 0.786) was statistically significant (p-value 0.005).
Vis fullstendig beskrivelse