engleski

Deep learning methods for segmentation of images of frozen tissue sections

There is an increasing demand for computer-aided diagnostic (CAD) systems to alleviate the pathologists’ workload. Due to limitation of diagnostic time that is of particular importance in intraoperative diagnostics. Thus, research described in this work attempts to implement such CAD system. The research is consisted of two preliminary studies and one full study. Preliminary studies are focused to the analysis of images of unstained and hematoxylin and eosin (H&E) stained frozen sections in order to diagnose adenocarcinoma of a colon in a liver using machine learning algorithms. The full study aims at creating a dataset with pixel-wise annotations and offers a baseline results for deep learning segmentation models. Unstained frozen sections were acquired using phase-contrast mode and preliminary results were presented for intraoperative image segmentation. Besides spectral angle mapper (SAM), two deep learning algorithms, U-Net and Structured autoencoder (SAE) for deep subspace clustering, were used to segment adenocarcinoma of a colon in a liver. SAM, U-Net and SAE were trained and tested on eighteen and eight images, respectively. Train set images belong to five unique patients while test images belong to three patients. In comparative analysis of performance metrics, the performance scores obtained by U-Net were the highest. Hence, yielded macro-averaged results were: balanced accuracy (BACC) of 83.70%±8%, sensitivity (SE) of 94.50% ±8%, specificity (SP) of 72.9%±8%, and F1 score of 45.20%±25.4%. Variations in the quality of phase- contrast images of unstained frozen sections are caused by several reasons including the absence of tissue fixation and ethanol-induced dehydration, thawing of the specimen under the microscope, and frozen crystals in the specimen. In consequence, this impacts the reproducibility of diagnostic performance. Regarding the preliminary work on hematoxylin and eosin (H&E) stained frozen sections, deep learning (DL) techniques appear to be well-suited for image analysis in CAD-based systems in digital pathology domain. Unfortunately, DL networks contain numerous parameters, requiring a large annotated training dataset. Thus, the lack of annotated images is likely the greatest obstacle in employing deep learning techniques in digital pathology domain. This is particularly true for intraoperative tissue examination of H&E frozen sections. To compensate for the lack of training images, a transfer learning strategy was employed on DL models’ training. Using a sliding window method, the trained model predicted pixels many times for distinct stride levels (contexts) for improved diagnostic performance. Threshold optimization with a BACC score demonstrated the effectiveness of such a technique, as this performance metric has increased significantly. Nested U-Net (U- Net++) model with DenseNet201 backbone significantly outperformed commonly used U-Net with VGG16 backbone on the preliminary dataset for both BACC and F1 score. Finally, in this research a public database was created to address the shortage of labeled histopathological images on a pixel-level. The dataset contains 82 histopathological images of H&E stained frozen sections obtained intraoperatively from 19 patients diagnosed with adenocarcinoma of a colon in a liver. Pixel-wise annotations by seven experts and the majority vote between all of them was also included. Experts were four pathologists, one senior student of medicine and two residents in pathology. Corresponding Fleiss’ kappa score of 0.74 shows substantial inter-annotator agreement. Moreover, besides originally stained images, the dataset consists of two stain-normalized versions of two sets of images stain-normalized with respect to two target images selected by the pathologist. Moreover, two conventional machine learning algorithms and three DL-based models were developed to provide baseline results. Specifically, incremental kNN, SVM, DeepLabv3+, UNet, and U-Net++ were used to perform binary (cancer vs. non-cancer) image segmentation. When combining results from originally stained and stain-normalized image sets, DL models achieve even better results. On an independent test set, deep learning models outperformed SVM and kNN by an average of 14% in terms of micro-averaged BACC, 15% in terms of micro-averaged F1 score, and 26% in terms of micro-averaged positive predicted value (PPV). On the other hand, performance scores of different DL models fall within 2%. The performance differences between the same DL architectures, with training from random initialization and from pretrained weights, were not statistically significant because of the domain gap. Finally, U-Net++ architecture with DenseNet201 backbone yielded the highest micro- averaged BACC score of 89.34%, F1 score of 83.67% and PPV of 81.11%. The database and tutorials are publicly available at https://cocahis.irb.hr.

intraoperatve diagnosis ; adenocarcinoma of a colon in a liver ; frozen sections ; light microscopy ; unstained specimens ; deep learning ; machine learning ; transfer learning ; segmentation ; hematoxylin-eosin ; stain-normalization

nije evidentirano