Tuomas Mirtti

Tolou Shadbahr, Michael Roberts, Jan Stanczuk et al.

Classifying samples in incomplete datasets is a common aim for machine learning practitioners, but is non-trivial. Missing data is found in most real-world datasets and these missing values are typically imputed using established methods, followed by classification of the now complete, imputed, samples. The focus of the machine learning researcher is then to optimise the downstream classification performance. In this study, we highlight that it is imperative to consider the quality of the imputation. We demonstrate how the commonly used measures for assessing quality are flawed and propose a new class of discrepancy scores which focus on how well the method recreates the overall distribution of the data. To conclude, we highlight the compromised interpretability of classifier models trained using poorly imputed data.

Joona Pohjonen, Carolin Stürenberg, Atte Föhr et al.

Deep neural networks have achieved impressive performance in a wide variety of medical imaging tasks. However, these models often fail on data not used during training, such as data originating from a different medical centre. How to recognize models suffering from this fragility, and how to design robust models are the main obstacles to clinical adoption. Here, we present general methods to identify causes for model generalisation failures and how to circumvent them. First, we use $\textit{distribution-shifted datasets}$ to show that models trained with current state-of-the-art methods are highly fragile to variability encountered in clinical practice, and then develop a $\textit{strong augmentation}$ strategy to address this fragility. Distribution-shifted datasets allow us to discover this fragility, which can otherwise remain undetected after validation against multiple external datasets. Strong augmentation allows us to train robust models achieving consistent performance under shifts from the training data distribution. Importantly, we demonstrate that strong augmentation yields biomedical imaging models which retain high performance when applied to real-world clinical data. Our results pave the way for the development and evaluation of reliable and robust neural networks in clinical practice.

Joona Pohjonen, Abderrahim-Oussama Batouche, Antti Rannikko et al.

Foundation models are trained on massive amounts of data to distinguish complex patterns and can be adapted to a wide range of downstream tasks with minimal computational resources. Here, we develop a foundation model for prostate cancer digital pathology called HistoEncoder by pre-training on 48 million prostate tissue tile images. We demonstrate that HistoEncoder features extracted from tile images with similar histological patterns map closely together in the feature space. HistoEncoder outperforms models pre-trained with natural images, even without fine-tuning or with 1000 times less training data. We describe two use cases that leverage the capabilities of HistoEncoder by fine-tuning the model with a limited amount of data and computational resources. First, we show how HistoEncoder can be used to automatically annotate large-scale datasets with high accuracy. Second, we combine histomics with commonly used clinical nomograms, significantly improving prostate cancer-specific death survival models. Foundation models such as HistoEncoder can allow organizations with limited resources to build effective clinical software tools without needing extensive datasets or significant amounts of computing.

Joona Pohjonen, Carolin Stürenberg, Antti Rannikko et al.

Many current neural networks for medical imaging generalise poorly to data unseen during training. Such behaviour can be caused by networks overfitting easy-to-learn, or statistically dominant, features while disregarding other potentially informative features. For example, indistinguishable differences in the sharpness of the images from two different scanners can degrade the performance of the network significantly. All neural networks intended for clinical practice need to be robust to variation in data caused by differences in imaging equipment, sample preparation and patient populations. To address these challenges, we evaluate the utility of spectral decoupling as an implicit bias mitigation method. Spectral decoupling encourages the neural network to learn more features by simply regularising the networks' unnormalised prediction scores with an L2 penalty, thus having no added computational costs. We show that spectral decoupling allows training neural networks on datasets with strong spurious correlations and increases networks' robustness for data distribution shifts. To validate our findings, we train networks with and without spectral decoupling to detect prostate cancer tissue slides and COVID-19 in chest radiographs. Networks trained with spectral decoupling achieve up to 9.5 percent point higher performance on external datasets. Our results show that spectral decoupling helps with generalisation issues associated with neural networks, and can be used to complement or replace computationally expensive explicit bias mitigation methods, such as stain normalization in histological images. We recommend using spectral decoupling as an implicit bias mitigation method in any neural network intended for clinical use.

Umair Akhtar Hasan Khan, Carolin Stürenberg, Oguzhan Gencoglu et al.

Deep neural networks have introduced significant advancements in the field of machine learning-based analysis of digital pathology images including prostate tissue images. With the help of transfer learning, classification and segmentation performance of neural network models have been further increased. However, due to the absence of large, extensively annotated, publicly available prostate histopathology datasets, several previous studies employ datasets from well-studied computer vision tasks such as ImageNet dataset. In this work, we propose a transfer learning scheme from breast histopathology images to improve prostate cancer detection performance. We validate our approach on annotated prostate whole slide images by using a publicly available breast histopathology dataset as pre-training. We show that the proposed cross-cancer approach outperforms transfer learning from ImageNet dataset.