The technique of whole slide imaging (WSI) boosts the application of deeplearning in medical imaging analysis and computational pathology. How-ever, fully supervised learning stucks into...Show moreThe technique of whole slide imaging (WSI) boosts the application of deeplearning in medical imaging analysis and computational pathology. How-ever, fully supervised learning stucks into bottlenecks due to the heavyreliance on manual annotations, which requires specific expertise and ex-pensive cost. Self-supervised learning would be a potential solution, whichis supervised by the signals generated from itself. It has been provedto perform as well as supervised learning on ImageNet in classificationtasks. Yet, its performance on medical image classification is unexplored.This study verifies the effectiveness of four self-supervised learning to de-tect anatomic structures on kidney biopsy WSI, including SimCLR, MoCo,SwAV and Barlow Twins. In the pretext-task, these self-supervised learn-ing algorithms are trained in 500 epochs with the same backbone archi-tecture, ResNet-50, which is initialized by the weights pre-trained on Im-ageNet correspondingly. The evaluation protocol is a semi-supervisedlinear classifier, implemented by using multi-nomial logistic regression.The results of the classification task show the features extracted by thefour algorithms all achieve good accuracy scores, higher than 85% withonly 10% labels. Among them, SwAV outperforms the other algorithmsfrom the perspective of overview and each class. Through this study,self-supervised learning algorithms exhibit the potential for more complextasks related to renal pathology.Show less
The number of mitotic cells in pathological sections is an important biomarker in the grad-ing and diagnosis of cancer and is used to assess tumor proliferation and aggressiveness.Detection of...Show moreThe number of mitotic cells in pathological sections is an important biomarker in the grad-ing and diagnosis of cancer and is used to assess tumor proliferation and aggressiveness.Detection of mitosis is primarily performed manually by trained pathologists examininghistopathology sections stained with hematoxylin and eosin (H&E) on glass slides througha microscope. This traditional approach is cumbersome, error-prone and has a high dis-crepancy between raters. With the advent of modern scanning devices, these specimenscan be digitised into whole slide images (WSI) and viewed and processed by a computer.Medical image processing algorithms can be applied to these images to more consistentlyand reliably assist doctors with many repetitive, tedious and time-consuming tasks. How-ever, there are large differences in tissue appearance between images from different scannervendors. This large variation results in a domain shift that limits the application of thesealgorithms in practice, as it can cause a large performance drop in algorithms applied tounseen images. Nevertheless, the domain dependent information is just an artefact of thescanning device and is often not relevant to the task at hand. In this work, we implementa method for automatic detection of mitotic figures using deep learning in H&E images ofbreast cancer. The method is developed using a collection of publicly available data andevaluated against a non-publicly available dataset from the biotechnology company Agen-dia in Amsterdam. The problem of domain shifts between images from different scanningdevices is addressed by implementing a technique from the domain adaptation literaturethat removes domain-dependent information from features during model training, result-ing in domain-invariant features that are only relevant to the task at hand and thereforeshould generalise well to unseen domains. We show that our model achieves competitiveperformance compared to results reported in the current literature and provide insightsinto how to assess the domain generalization capability of the models. The evaluationon the Agendia dataset shows the effects of inter-rater variability and highlights the im-portance of creating high quality medical datasets for the development and evaluationof medical image processing algorithms. Further experiments are necessary to evaluatethe domain generalization of the model, however the results represent a promising steptowards more robust algorithms for the application in clinical practice.Show less