Conditional Deep Convolutional Neural Networks for Improved Generalization of Automated Screening of Histopathological Images

The increase in computing power of the last two decades has fueled the growth of a new field of pathology, digital pathology, where glass slides are digitized with high-speed scanners that produce multi-gigabyte images, Whole Slide Images (WSIs). Since the advent of AlexNet in 2012, multiple works have successfully applied deep learning based computer vision to histopathology with performances comparable to the ones of human experts. However, such algorithms have seen limited adoption in the clinical practice of histopathology for two main reasons: - deep learning algorithms, trained on datasets of WSIs collected from one or more medical centers, give very accurate results, e.g. classifications of specimen as healthy or diseased, when tested on WSIs from the same centers, but tend to perform significantly worse when tested on datasets of WSIs acquired by different medical centers; - deep learning algorithms show limited interpretability of the results, i.e. they offer limited insights into how and why particular results were obtained. In digital pathology, the first concern is significant because there exist large variations in the data characteristics of histopathology datasets acquired by different medical centers, this is known as "domain shift". Furthermore, the number of curated, publicly available datasets, to be used for training is limited; as such improving the performance of deep learning algorithms through data volume and standard training techniques is not a viable solution to the domain shift problem. In this thesis, I focus on studying and implementing a method drawn from the "few-shot learning" paradigm, an approach to train and test deep learning algorithms with few examples, to be robust to domain shift. To address the second concern of limited interpretability, I implemented a Fully Convolutional Network (FCN), adapted to few-shot learning, for automatic segmentation of metastases in WSIs. Because FCNs output a lesion probability heatmap which can be overlaid on top of the input WSI, the interpretability of algorithmic decisions is easier as it can be related to hue and pattern appearances of the underlying image. The goal is to provide a decision support tool that could assist the pathologists in screening WSIs and that could highlight areas where their evaluation is needed. To achieve this goal, I focus on histopathology images of sentinel lymph nodes for the diagnosis of breast cancer. WSIs in the dataset were fully annotated by expert pathologists to contour metastases. For this research, I selected and studied a FCN based algorithms whose final predictions can be guided at inference time by providing as input, together with the WSI to segment, a set of other images, known as "support" images, that condition the final output. The architecture is a variant of conditional FCNs (co-FCNs). I also identified a protocol, that partially relies on unsupervised learning techniques, to associate each input WSI to the appropriate support images. Finally, I conducted experiments to evaluate the performance of such algorithm and I compared its performance against a standard FCN semantic segmentation algorithm. My main contributions are: - I have studied a few-shot learning method to address the issue of domain shift in digital pathology; - I have identified and implemented the necessary architectural changes to make the chosen co-FCN architecture applicable to segment WSIs; - I have devised and implemented a method for the selection of the support set necessary to enable effective few-shot learning . This is also, to the best of my knowledge, the first study of the applicability of co-FCNs to digital pathology.The increase in computing power of the last two decades has fueled the growth of a new field of pathology, digital pathology, where glass slides are digitized with high-speed scanners that produce multi-gigabyte images, Whole Slide Images (WSIs). Since the advent of AlexNet in 2012, multiple works have successfully applied deep learning based computer vision to histopathology with performances comparable to the ones of human experts. However, such algorithms have seen limited adoption in the clinical practice of histopathology for two main reasons: - deep learning algorithms, trained on datasets of WSIs collected from one or more medical centers, give very accurate results, e.g. classifications of specimen as healthy or diseased, when tested on WSIs from the same centers, but tend to perform significantly worse when tested on datasets of WSIs acquired by different medical centers; - deep learning algorithms show limited interpretability of the results, i.e. they offer limited insights into how and why particular results were obtained. In digital pathology, the first concern is significant because there exist large variations in the data characteristics of histopathology datasets acquired by different medical centers, this is known as "domain shift". Furthermore, the number of curated, publicly available datasets, to be used for training is limited; as such improving the performance of deep learning algorithms through data volume and standard training techniques is not a viable solution to the domain shift problem. In this thesis, I focus on studying and implementing a method drawn from the "few-shot learning" paradigm, an approach to train and test deep learning algorithms with few examples, to be robust to domain shift. To address the second concern of limited interpretability, I implemented a Fully Convolutional Network (FCN), adapted to few-shot learning, for automatic segmentation of metastases in WSIs. Because FCNs output a lesion probability heatmap which can be overlaid on top of the input WSI, the interpretability of algorithmic decisions is easier as it can be related to hue and pattern appearances of the underlying image. The goal is to provide a decision support tool that could assist the pathologists in screening WSIs and that could highlight areas where their evaluation is needed. To achieve this goal, I focus on histopathology images of sentinel lymph nodes for the diagnosis of breast cancer. WSIs in the dataset were fully annotated by expert pathologists to contour metastases. For this research, I selected and studied a FCN based algorithms whose final predictions can be guided at inference time by providing as input, together with the WSI to segment, a set of other images, known as "support" images, that condition the final output. The architecture is a variant of conditional FCNs (co-FCNs). I also identified a protocol, that partially relies on unsupervised learning techniques, to associate each input WSI to the appropriate support images. Finally, I conducted experiments to evaluate the performance of such algorithm and I compared its performance against a standard FCN semantic segmentation algorithm. My main contributions are: - I have studied a few-shot learning method to address the issue of domain shift in digital pathology; - I have identified and implemented the necessary architectural changes to make the chosen co-FCN architecture applicable to segment WSIs; - I have devised and implemented a method for the selection of the support set necessary to enable effective few-shot learning . This is also, to the best of my knowledge, the first study of the applicability of co-FCNs to digital pathology