The Emerging Issues in Bioimaging AI Publications and Research

Deep learning (DL) is now a powerful tool in microscopy data analysis, routinely used for image processing applications such as segmentation and denoising. However, it is rarely used to directly learn scientific models of a biological system, owing to the complexity of the internal representations. Here, we present our recent attempts to learn interpretable DL-based models of complex cell biological phenomena directly from a large corpus of time-lapse imaging data. In particular, we implement disentangled representation learning, causal time series models, network sparsity and symbolic methods, and assess their usefulness in forming interpretable models of complex data. We find that such methods can produce highly parsimonious models that achieve $\sim 98\%$ of the accuracy of black-box benchmark models, with a tiny fraction of the complexity. We explore the utility of such interpretable models in producing scientific explanations of the underlying biological phenomenon.

The black box nature of neural networks is commonly regarded as the main source why predictions obtained from deep neural networks, despite their often unpredecedented predictive accuracy, are often considered untrustworthy. In this contribution, I argue that the lack of trustworthiness of machine learning in general is due to its inductive nature: Machine learning models are obtained from inductive inferences, where specific observations in the form of training data are used to infer a general model that can classify data points beyond the training data. From this perspective, machine learning is subject to the problem of induction, which has been brought to the point by the no-free-lunch theorem: Since there is no justification to assume that future events will resemble the past, all machine learning algorithms perform equal in terms of their out-of-training error.

Our further reasoning follows two interpretations, a global and a local interpretation, of the no-free-lunch theorem, which have been formulated recently by Sterkenburg and Grünwald. The global interpretation is in a sense the pessimistic interpretation, stating that no universal learning algorithm exists, since across the domain of all possible learning problems, all classifiers are identical in terms of their out-of-training error. The local interpretation is more constructive towards applied machine learning: When dealing with one specific problem, some learning algorithms do perform better on this specific task than other learning algorithms. This can be understood in terms of the inductive bias of different learning algorithms: As a direct implication of the no-free-lunch theorem, each learning algorithms must involve an either implicit or an explicit set of assumptions about how to generalize to data points beyond the training data. This set of assumptions is referred to as the inductive bias of a learning algorithm. From the perspective of one specific learning task, one can now ask what learning algorithm has an inductive bias that matches the underlying learning problem. This local interpretation of the no-free-lunch theorem essentially leads to considering machine learning as an inductive bias modeling problem.

The question that follows the local interpretation of the no-free-lunch theorem is how to justify inductive bias. I argue that our recently proposed framework for falsifiable explanations of artificial intelligence, or FXAI framework for short, addresses this question: The FXAI framework builds on the concept of explainability methods for neural networks, which usually provide an explainable output along with the classification of an input item. In the case of image classification, for example, the interepretable output is often a heat map that indicates which input variables have been relevant for obtaining the classification result of a specific image. In the FXAI framework, this explainable extension of the output is referred to as the interepretable space, or I-space for short. It is important to realize that an I-space, while being interpretable, can usually not be considered an interpretation in itself. The role of an interpretation (or, synonymously, an explanation) is rather assigned to a hypothesis that, in the sense of a scientific hypothesis, is required to be experimentally testable. The latter criterion is of crucial importance: Now, the explanation – and along with it, the I-space and hence the machine learning model – can be tested experimentally.

Experimental testability has relevant consequences: First of all, since the experiment that tests the explaining hypothesis is a different experiment than the experiment that yielded the data that were input to the machine learning model, the FXAI framework yields an experimental, deductive path to validate a machine learning model that is fully independent of cross validation. Second, the testable hypothesis suggests what should guide the inductive bias of a learning algorithm: namely an experimentally testable hypothesis.

We can now finally argue why experimentally testable explanations improve the trustworthiness of machine learning models. My argument lies in the nature of scientific hypotheses, which usually do not refer to one specific experiment. Rather, a strong hypothesis will usually suggest a wide range of different experiments through which the hypothesis can be tested. The more experiments a hypothesis invites for it to be tested, the more vulnerable the hypothesis becomes, because each experiment potentially falsifies the hypothesis. If, on the other hand, the hypothesis withstands all experimental attempts to falsify it, then the trustworthiness of the hypothesis and with it the associated machine learning model is undermined.

The segmentation of cells in light microscopy or organelles in electron microscopy is one of the fundamental tasks in microscopy image analysis. While deep learning based approaches have improved segmentation qualities for a wide array of tasks, these solutions require specialized architectures and, unless very similar training data is publicly available, a significant amount of manual annotation. Recently versatile models that can be applied to a wider set of vision tasks – commonly referred to as foundation models – have been introduced. These models promise to bridge this gap and enable readily available solutions for many vision tasks. The foundation model “Segment Anything” developed by Meta implements this paradigm for segmentation tasks and can be applied for interactive and automatic segmentation in a large variety of image modalities. Our work builds on Segment Anything and evaluates and improves it for microscopy data. In particular, we implement a fine-tuning methodology that significantly improves the quality for microscopy and a software plugin for fast interactive data annotation., showing the promise of vision foundation models for microscopy image analysis.

To facilitate the characterization of unlabeled induced pluripotent stem cells (iPSCs) during culture and expansion, and to be able to address gene expression in individual living cells over time, we developed an AI pipeline for nuclear segmentation and mitosis detection from phase contrast images of individual cells within iPSC colonies. The analysis uses a 2D convolutional neural network (U-Net) plus a 3D U-Net applied on time lapse images to detect and segment nuclei, mitotic events, and daughter nuclei to enable tracking of hundreds of thousands of individual cells over long times in culture. The analysis uses fluorescence data to train models for segmenting nuclei in phase contrast images. The use of classical image processing routines to segment fluorescent nuclei precludes the need for manual annotation and provides hundreds of thousands of cell objects for training. We explored reproducibility and generalizability of the pipeline, and how pipeline parameters influenced metrics of accuracy. The model is generalizable in that it performs well on different datasets with an average F1 score of 0.94, on cells at different densities, and on cells from different pluripotent cell lines. The method allows us to assess, in a non-invasive manner, rates of mitosis and cell division which serve as indicators of cell state and cell health. We assess these parameters in culture for more than 36 hours, at different locations in the colonies, and as a function of excitation light exposure.

Reporting microscopy data and metadata are critical for data reproducibility, sharing, and reuse, and journals can have a key role in improving reporting standards. This talk discussed a methodological reporting crisis in microscopy, published works seeking to address this issue, and standards that are being implemented at Nature Methods. It also discussed the unique challenges associated with publishing reproducible AI for use in bioimage analysis and why this is crucial for the field moving forward. The goal was to inspire researchers to develop and implement best practice to promote reproducibility and growth within the field.

Neural networks trained through optimization techniques based on variants of stochastic gradient descent (SGD) present a spectral bias. Model training dynamics are biased towards learning features related to low-frequency components of the input data at early stages of training, and subsequently focusing on high-frequency features. Another important phenomenon in training dynamics is the emergence of shortcut learning, that is learning spurious correlations between the input data and prediction target. This results from the tendency of SGD-based training to find solutions that simplify the minimization of a loss function used as target of the training optimization problem. Shortcuts harm the generalization abilities of neural networks, especially in out-of-distribution (OOD) settings, and thus require particular attention during model validation.

We investigate the relationship between spectral bias and shortcut learning in image classification and expose the existence of shortcuts learned by vision models in the Fourier domain, which we call frequency shortcuts. We propose a method to detect possible frequency shortcuts, based on the importance that single frequency components have in the classification task, and construct dominant frequency maps (DFM). We demonstrate that frequency shortcuts can be learned at low or high-frequency and potentially harm the generalization capabilities in out-of-distribution settings, showing that shortcuts presents in OOD data can cause an illusion of strong generalization. In order to mitigate their impact on model performance, we also investigate the use of DFMs in a negative data augmentation strategy that improves adversarial robustness. However, extensive analysis of shortcuts learned by vision models is necessary and requires substantial attention to validate model performance and transferability to real-world tasks.

With the rapid growing volume and complexity of modern biomedical visual data, we can no longer rely on humans’ amazing capacity to identify visual patterns in biomedical images. Deep learning has emerged as a powerful technique to identify hidden patterns that exceed human intuition in complex cell imaging data. Extracting a deeper biological understanding, such as mechanistic description of complex phenotypes, require human interpretable explanation of the deep learning model’s decision process, however, the non-linear entanglement of image features makes deep learning models a “black box” that lacks straightforward explanations of which biologically meaningful image properties are important for the models’ decision. In my talk I presented a new generalized method toward systematic visual interpretability of deep learning image-based classification models that relies on counterfactual visual explanations using a disentangled latent representation. This method enables visually intuitive traversal of the latent space and we applied it to decipher blastocysts morphological quality properties in the context of in vitro fertilization.

In this presentation, we delve into the potential of Foundation Models (FMs), which encompass Large Language Models (LLMs), Large Vision Models (VLMs), and Multimodal Large Language Models (MLLM). These models have demonstrated promising outcomes in various scenarios. However, integrating FM capabilities into professional domains remains an unresolved inquiry. We present some recent relevant research endeavours to address emergent challenges during this transition. The initial query pertains to efficiently aligning data from specialized domains with non-specific FMs. Parameter Efficient Fine-tuning (PEFT) offers a viable approach, and to adapt PEFT more suitably for medical data in our case, we have devised a tree-like structured adapter that hierarchically incorporates medical knowledge into the Segment Anything (SAM) model. Secondly, we illustrate how quantization techniques can accelerate FM performance for biological tasks. Subsequently, we demonstrate the fine-tuning process of an MLLM using medical data to generate medical reports and accomplish vision-based question-answering tasks. This process leverages methods such as in-context learning to align training data across different modalities with retrieval augmentative generation to support our model giving a more comprehensive report.

In this talk, I will discuss two important aspects of machine learning: algorithmic fairness and robust generalization under the common framework of invariant causal prediction. I will first provide some motivating examples of these two problems in the context of biomedical and healthcare applications. I will then introduce our recent work [1, 2] on invariant feature recovery to address the above two problems. I will conclude the talk with a discussion of some open problems and future research directions.