Sybren Meijer - U-Net ensembles for accurate classification of esophageal adenocarcinoma
Background & objectives
Esophageal adenocarcinoma has a dismal prognosis and Barrett's esophagus (BE) is the only known precursor lesion. BE progresses through a metaplasia-dysplasia-carcinoma sequence. Progression rates from non-dysplastic BE (NDBE) are low (0,5%), but a histopathological diagnosis of low grade dysplasia (LGD) is a strong independent risk factor for progression. As a result of significant interobserver variation, reported progression rates vary from 1- 40% and therefore International guidelines mandate a second opinion by an expert-pathologist. We aim to develop a convolutional neural network for objective and reproducible diagnosis of dysplasia in BE at expert-pathologist level.
TIFF images from 400 digitized biopsies of 170 BE patients were annotated in high detail by an expert pathologist, generating binary classification masks for NDBE and dysplastic glands. Patches of the H&E staining and corresponding binary mask were extracted, resulting in 148.033 and 34.557 patches for training and testing, respectively. An ensemble of U-Nets was used, using a combination of the U-Net architecture with DenseNet and ResNet models, pre-trained on ImageNet. The ensemble consisted of four U-Nets with a down-sampling path using the DenseNet architecture and four other U-nets using the ResNet architecture. As metrics F1 coefficient and the pixel-wise accuracy was used.
F1 scores (range 80.6 and 84.9%) and pixel-wise accuracy (range 86.5 to 90.6%) of individual models and combinations; the best performing combination was the ensemble consisting of a committee of eight DenseNet and ResNet models.
Model F1-score Pixel-wise accuracy
U-Net 80,6% 86,5%
U-Net + ResNet-34 81.9 % 87.6 %
U-Net + ResNet-152 83.1 % 88.9 %
U-Net + DenseNet-161 82.8 % 88.1 %
Ensemble DenseNet and ResNet 84.9 % 90.6 %
U-Net ensembles that classify precursor lesions of esophageal adenocarcinoma can be used for accurate risk stratification of patients with BE.