Federated learning for computational pathology on gigapixel whole slide images

Deep Learning-based computational pathology algorithms have demonstrated profound ability to excel in a wide array of tasks that range from characterization of well known morphological phenotypes to predicting non human-identifiable features from histology such as molecular alterations. However, the development of robust, adaptable and accurate deep learning-based models often rely on the collection and time-costly curation large high-quality annotated training data that should ideally come from diverse sources and patient populations to cater for the heterogeneity that exists in such datasets. Multi-centric and collaborative integration of medical data across multiple institutions can naturally help overcome this challenge and boost the model performance but is limited by privacy concerns among other difficulties that may arise in the complex data sharing process as models scale towards using hundreds of thousands of gigapixel whole slide images. In this paper, we introduce privacy-preserving federated learning for gigapixel whole slide images in computational pathology using weakly-supervised attention multiple instance learning and differential privacy. We evaluated our approach on two different diagnostic problems using thousands of histology whole slide images with only slide-level labels. Additionally, we present a weakly-supervised learning framework for survival prediction and patient stratification from whole slide images and demonstrate its effectiveness in a federated setting. Our results show that using federated learning, we can effectively develop accurate weakly-supervised deep learning models from distributed data silos without direct data sharing and its associated complexities, while also preserving differential privacy using randomized noise generation. We also make available an easy-to-use federated learning for computational pathology software package: http://github.com/mahmoodlab/HistoFL.     

Weakly supervised instance learning for thyroid malignancy prediction from whole slide cytopathology images

We consider machine-learning-based thyroid-malignancy prediction from cytopathology whole-slide images (WSI). Multiple instance learning (MIL) approaches, typically used for the analysis of WSIs, divide the image (bag) into patches (instances), which are used to predict a single bag-level label. These approaches perform poorly in cytopathology slides due to a unique bag structure: sparsely located informative instances with varying characteristics of abnormality. We address these challenges by considering multiple types of labels: bag-level malignancy and ordered diagnostic scores, as well as instance-level informativeness and abnormality labels. We study their contribution beyond the MIL setting by proposing a maximum likelihood estimation (MLE) framework, from which we derive a two-stage deep-learning-based algorithm. The algorithm identifies informative instances and assigns them local malignancy scores that are incorporated into a global malignancy prediction. We derive a lower bound of the MLE, leading to an improved training strategy based on weak supervision, that we motivate through statistical analysis. The lower bound further allows us to extend the proposed algorithm to simultaneously predict multiple bag and instance-level labels from a single output of a neural network. Experimental results demonstrate that the proposed algorithm provides competitive performance compared to several competing methods, achieves (expert) human-level performance, and allows augmentation of human decisions.     

A hybrid network for automatic hepatocellular carcinoma segmentation in H&E-stained whole slide images

Hepatocellular carcinoma (HCC), as the most common type of primary malignant liver cancer, has become a leading cause of cancer deaths in recent years. Accurate segmentation of HCC lesions is critical for tumor load assessment, surgery planning, and postoperative examination. As the appearance of HCC lesions varies greatly across patients, traditional manual segmentation is a very tedious and time-consuming process, the accuracy of which is also difficult to ensure. Therefore, a fully automated and reliable HCC segmentation system is in high demand. In this work, we present a novel hybrid neural network based on multi-task learning and ensemble learning techniques for accurate HCC segmentation of hematoxylin and eosin (H&E)-stained whole slide images (WSIs). First, three task-specific branches are integrated to enlarge the feature space, based on which the network is able to learn more general features and thus reduce the risk of overfitting. Second, an ensemble learning scheme is leveraged to perform feature aggregation, in which selective kernel modules (SKMs) and spatial and channel-wise squeeze-and-excitation modules (scSEMs) are adopted for capturing the features from different spaces and scales. Our proposed method achieves state-of-the-art performance on three publicly available datasets, with segmentation accuracies of 0.797, 0.923, and 0.765 in the PAIP, CRAG, and UHCMC&CWRU datasets, respectively, which demonstrates its effectiveness in addressing the HCC segmentation problem. To the best of our knowledge, this is also the first work on the pixel-wise HCC segmentation of H&E-stained WSIs.     

Conditional GANs based system for fibrosis detection and quantification in Hematoxylin and Eosin whole slide images

Assessing the degree of liver fibrosis is fundamental for the management of patients with chronic liver disease, in liver transplants procedures, and in general liver disease research. The fibrosis stage is best assessed by histopathologic evaluation, and Masson’s Trichrome stain (MT) is the stain of choice for this task in many laboratories around the world. However, the most used stain in histopathology is Hematoxylin Eosin (HE) which is cheaper, has a faster turn-around time and is the primary stain routinely used for evaluation of liver specimens. In this paper, we propose a novel digital pathology system that accurately detects and quantifies the footprint of fibrous tissue in HE whole slide images (WSI). The proposed system produces virtual MT images from HE using a deep learning model that learns deep texture patterns associated with collagen fibers. The training pipeline is based on conditional generative adversarial networks (cGAN), which can achieve accurate pixel-level transformation. Our comprehensive training pipeline features an automatic WSI registration algorithm, which qualifies the HE/MT training slides for the cGAN model. Using liver specimens collected during liver transplantation procedures, we conducted a range of experiments to evaluate the detected footprint of selected anatomical features. Our evaluation includes both image similarity and semantic segmentation metrics. The proposed system achieved enhanced results in the experiments with significant improvement over the state-of-the-art CycleGAN learning style, and over direct prediction of fibrosis in HE without having the virtual MT step.     

Handcrafted Histological Transformer (H2T): Unsupervised representation of whole slide images

Diagnostic, prognostic and therapeutic decision-making of cancer in pathology clinics can now be carried out based on analysis of multi-gigapixel tissue images, also known as whole-slide images (WSIs). Recently, deep convolutional neural networks (CNNs) have been proposed to derive unsupervised WSI representations; these are attractive as they rely less on expert annotation which is cumbersome. However, a major trade-off is that higher predictive power generally comes at the cost of interpretability, posing a challenge to their clinical use where transparency in decision-making is generally expected. To address this challenge, we present a handcrafted framework based on deep CNN for constructing holistic WSI-level representations. Building on recent findings about the internal working of the Transformer in the domain of natural language processing, we break down its processes and handcraft them into a more transparent framework that we term as the Handcrafted Histological Transformer or H2T. Based on our experiments involving various datasets consisting of a total of 10,042 WSIs, the results demonstrate that H2T based holistic WSI-level representations offer competitive performance compared to recent state-of-the-art methods and can be readily utilized for various downstream analysis tasks. Finally, our results demonstrate that the H2T framework can be up to 14 times faster than the Transformer models.     

Supervised graph hashing for histopathology image retrieval and classification

In pathology image analysis, morphological characteristics of cells are critical to grade many diseases. With the development of cell detection and segmentation techniques, it is possible to extract cell-level information for further analysis in pathology images. However, it is challenging to conduct efficient analysis of cell-level information on a large-scale image dataset because each image usually contains hundreds or thousands of cells. In this paper, we propose a novel image retrieval based framework for large-scale pathology image analysis. For each image, we encode each cell into binary codes to generate image representation using a novel graph based hashing model and then conduct image retrieval by applying a group-to-group matching method to similarity measurement. In order to improve both computational efficiency and memory requirement, we further introduce matrix factorization into the hashing model for scalable image retrieval. The proposed framework is extensively validated with thousands of lung cancer images, and it achieves 97.98% classification accuracy and 97.50% retrieval precision with all cells of each query image used.     

Benchmarking weakly-supervised deep learning pipelines for whole slide classification in computational pathology

Artificial intelligence (AI) can extract visual information from histopathological slides and yield biological insight and clinical biomarkers. Whole slide images are cut into thousands of tiles and classification problems are often weakly-supervised: the ground truth is only known for the slide, not for every single tile. In classical weakly-supervised analysis pipelines, all tiles inherit the slide label while in multiple-instance learning (MIL), only bags of tiles inherit the label. However, it is still unclear how these widely used but markedly different approaches perform relative to each other.We implemented and systematically compared six methods in six clinically relevant end-to-end prediction tasks using data from N=2980 patients for training with rigorous external validation. We tested three classical weakly-supervised approaches with convolutional neural networks and vision transformers (ViT) and three MIL-based approaches with and without an additional attention module. Our results empirically demonstrate that histological tumor subtyping of renal cell carcinoma is an easy task in which all approaches achieve an area under the receiver operating curve (AUROC) of above 0.9. In contrast, we report significant performance differences for clinically relevant tasks of mutation prediction in colorectal, gastric, and bladder cancer. In these mutation prediction tasks, classical weakly-supervised workflows outperformed MIL-based weakly-supervised methods for mutation prediction, which is surprising given their simplicity. This shows that new end-to-end image analysis pipelines in computational pathology should be compared to classical weakly-supervised methods. Also, these findings motivate the development of new methods which combine the elegant assumptions of MIL with the empirically observed higher performance of classical weakly-supervised approaches. We make all source codes publicly available at https://github.com/KatherLab/HIA, allowing easy application of all methods to any similar task.     

InstantScope: a low-cost whole slide imaging system with instant focal plane detection

We report the development of a high-throughput whole slide imaging (WSI) system by adapting a cost-effective optomechanical add-on kit to existing microscopes. Inspired by the phase detection concept in professional photography, we attached two pinhole-modulated cameras at the eyepiece ports for instant focal plane detection. By adjusting the positions of the pinholes, we can effectively change the view angle for the sample, and as such, we can use the translation shift of the two pinhole-modulated images to identify the optimal focal position. By using a small pinhole size, the focal-plane-detection range is on the order of millimeter, orders of magnitude longer than the objective’s depth of field. We also show that, by analyzing the phase correlation of the pinhole-modulated images, we can determine whether the sample contains one thin section, folded sections, or multiple layers separated by certain distances – an important piece of information prior to a detailed z scan. In order to achieve system automation, we deployed a low-cost programmable robotic arm to perform sample loading and $14 stepper motors to drive the microscope stage to perform x-y scanning. Using a 20X objective lens, we can acquire a 2 gigapixel image with 14 mm by 8 mm field of view in 90 seconds. The reported platform may find applications in biomedical research, telemedicine, and digital pathology. It may also provide new insights for the development of high-content screening instruments.OCIS codes: (170.0180) Microscopy, (170.3010) Image reconstruction techniques, (110.0110) Imaging systems, (110.1220) Apertures     

Dual-path network with synergistic grouping loss and evidence driven risk stratification for whole slide cervical image analysis

Cervical cancer has been one of the most lethal cancers threatening women’s health. Nevertheless, the incidence of cervical cancer can be effectively minimized with preventive clinical management strategies, including vaccines and regular screening examinations. Screening cervical smears under microscope by cytologist is a widely used routine in regular examination, which consumes cytologists’ large amount of time and labour. Computerized cytology analysis appropriately caters to such an imperative need, which alleviates cytologists’ workload and reduce potential misdiagnosis rate. However, automatic analysis of cervical smear via digitalized whole slide images (WSIs) remains a challenging problem, due to the extreme huge image resolution, existence of tiny lesions, noisy dataset and intricate clinical definition of classes with fuzzy boundaries. In this paper, we design an efficient deep convolutional neural network (CNN) with dual-path (DP) encoder for lesion retrieval, which ensures the inference efficiency and the sensitivity on both tiny and large lesions. Incorporated with synergistic grouping loss (SGL), the network can be effectively trained on noisy dataset with fuzzy inter-class boundaries. Inspired by the clinical diagnostic criteria from the cytologists, a novel smear-level classifier, i.e., rule-based risk stratification (RRS), is proposed for accurate smear-level classification and risk stratification, which aligns reasonably with intricate cytological definition of the classes. Extensive experiments on the largest dataset including 19,303 WSIs from multiple medical centers validate the robustness of our method. With high sensitivity of 0.907 and specificity of 0.80 being achieved, our method manifests the potential to reduce the workload for cytologists in the routine practice.     

A graph-based approach for the retrieval of multi-modality medical images

In this paper, we address the retrieval of multi-modality medical volumes, which consist of two different imaging modalities, acquired sequentially, from the same scanner. One such example, positron emission tomography and computed tomography (PET-CT), provides physicians with complementary functional and anatomical features as well as spatial relationships and has led to improved cancer diagnosis, localisation, and staging.The challenge of multi-modality volume retrieval for cancer patients lies in representing the complementary geometric and topologic attributes between tumours and organs. These attributes and relationships, which are used for tumour staging and classification, can be formulated as a graph. It has been demonstrated that graph-based methods have high accuracy for retrieval by spatial similarity. However, naïvely representing all relationships on a complete graph obscures the structure of the tumour-anatomy relationships.We propose a new graph structure derived from complete graphs that structurally constrains the edges connected to tumour vertices based upon the spatial proximity of tumours and organs. This enables retrieval on the basis of tumour localisation. We also present a similarity matching algorithm that accounts for different feature sets for graph elements from different imaging modalities. Our method emphasises the relationships between a tumour and related organs, while still modelling patient-specific anatomical variations. Constraining tumours to related anatomical structures improves the discrimination potential of graphs, making it easier to retrieve similar images based on tumour location.We evaluated our retrieval methodology on a dataset of clinical PET-CT volumes. Our results showed that our method enabled the retrieval of multi-modality images using spatial features. Our graph-based retrieval algorithm achieved a higher precision than several other retrieval techniques: gray-level histograms as well as state-of-the-art methods such as visual words using the scale- invariant feature transform (SIFT) and relational matrices representing the spatial arrangements of objects.     

Optimum slicing of radical prostatectomy specimens for correlation between histopathology and medical images

Purpose  

There is a need for methods which enable precise correlation of histologic sections with in vivo prostate images. Such methods would allow direct comparison between imaging features and functional or histopathological heterogeneity of tumors. Correlation would be particularly useful for validating the accuracy of imaging modalities, developing imaging techniques, assessing image-guided therapy, etc. An optimum prostate slicing method for accurate correlation between the histopathological and medical imaging planes in terms of section angle, thickness and level was sought.     

Low-cost whole slide imaging system with single-shot autofocusing based on color-multiplexed illumination and deep learning

Recent research on whole slide imaging (WSI) has greatly promoted the development of digital pathology. However, accurate autofocusing is still the main challenge for WSI acquisition and automated digital microscope. To address this problem, this paper describes a low cost WSI system and proposes a fast, robust autofocusing method based on deep learning. We use a programmable LED array for sample illumination. Before the brightfield image acquisition, we turn on a red and a green LED, and capture a color-multiplexed image, which is fed into a neural network for defocus distance estimation. After the focus tracking process, we employ a low-cost DIY adaptor to digitally adjust the photographic lens instead of the mechanical stage to perform axial position adjustment, and acquire the in-focus image under brightfield illumination. To ensure the calculation speed and image quality, we build a network model based on a ‘light weight’ backbone network architecture-MobileNetV3. Since the color-multiplexed coherent illuminated images contain abundant information about the defocus orientation, the proposed method enables high performance of autofocusing. Experimental results show that the proposed method can accurately predict the defocus distance of various types of samples and has good generalization ability for new types of samples. In the case of using GPU, the processing time for autofocusing is less than 0.1 second for each field of view, indicating that our method can further speed up the acquisition of whole slide images.     

HookNet: Multi-resolution convolutional neural networks for semantic segmentation in histopathology whole-slide images

We propose HookNet, a semantic segmentation model for histopathology whole-slide images, which combines context and details via multiple branches of encoder-decoder convolutional neural networks. Concentric patches at multiple resolutions with different fields of view, feed different branches of HookNet, and intermediate representations are combined via a hooking mechanism. We describe a framework to design and train HookNet for achieving high-resolution semantic segmentation and introduce constraints to guarantee pixel-wise alignment in feature maps during hooking. We show the advantages of using HookNet in two histopathology image segmentation tasks where tissue type prediction accuracy strongly depends on contextual information, namely (1) multi-class tissue segmentation in breast cancer and, (2) segmentation of tertiary lymphoid structures and germinal centers in lung cancer. We show the superiority of HookNet when compared with single-resolution U-Net models working at different resolutions as well as with a recently published multi-resolution model for histopathology image segmentation. We have made HookNet publicly available by releasing the source code¹ as well as in the form of web-based applications^2,3 based on the grand-challenge.org platform.     

Multi-channel auto-encoders for learning domain invariant representations enabling superior classification of histopathology images

Domain shift is a problem commonly encountered when developing automated histopathology pipelines. The performance of machine learning models such as convolutional neural networks within automated histopathology pipelines is often diminished when applying them to novel data domains due to factors arising from differing staining and scanning protocols. The Dual-Channel Auto-Encoder (DCAE) model was previously shown to produce feature representations that are less sensitive to appearance variation introduced by different digital slide scanners. In this work, the Multi-Channel Auto-Encoder (MCAE) model is presented as an extension to DCAE which learns from more than two domains of data. Experimental results show that the MCAE model produces feature representations that are less sensitive to inter-domain variations than the comparative StaNoSA method when tested on a novel synthetic dataset. This was apparent when applying the MCAE, DCAE, and StaNoSA models to three different classification tasks from unseen domains. The results of this experiment show the MCAE model out performs the other models. These results show that the MCAE model is able to generalise better to novel data, including data from unseen domains, than existing approaches by actively learning normalised feature representations.     

Single image super-resolution for whole slide image using convolutional neural networks and self-supervised color normalization

High-quality whole slide scanners used for animal and human pathology scanning are expensive and can produce massive datasets, which limits the access to and adoption of this technique. As a potential solution to these challenges, we present a deep learning-based approach making use of single image super-resolution (SISR) to reconstruct high-resolution histology images from low-resolution inputs. Such low-resolution images can easily be shared, require less storage, and can be acquired quickly using widely available low-cost slide scanners. The network consists of multi-scale fully convolutional networks capable of capturing hierarchical features. Conditional generative adversarial loss is incorporated to penalize blurriness in the output images. The network is trained using a progressive strategy where the scaling factor is sampled from a normal distribution with an increasing mean. The results are evaluated with quantitative metrics and are used in a clinical histopathology diagnosis procedure which shows that the SISR framework can be used to reconstruct high-resolution images with clinical level quality. We further propose a self-supervised color normalization method that can remove staining variation artifacts. Quantitative evaluations show that the SISR framework can generalize well on unseen data collected from other patient tissue cohorts by incorporating the color normalization method.     

Lightsheet-based flow cytometer for whole blood with the ability for the magnetic retrieval of objects from the blood flow

Detection and extraction of circulating tumor cells and other rare objects in the bloodstream are of great interest for modern diagnostics, but devices that can solve this problem for the whole blood volume of laboratory animals are still rare. Here we have developed SPIM-based lightsheet flow cytometer for the detection of fluorescently-labeled objects in whole blood. The bypass channel between two blood vessels connected with the external flow cell was used to visualize, detect, and magnetically separate fluorescently-labeled objects without hydrodynamic focusing. Carriers for targeted drug delivery were used as model objects to test the device performance. They were injected into the bloodstream of the rat, detected fluorescently, and then captured from the bloodstream by a magnetic separator prior to filtration in organs. Carriers extracted from the whole blood were studied by a number of in vitro methods.