Vision Transformers (ViTs), such as DINOv2, achieve strong performance across domains but often repurpose low-informative patch tokens in ways that reduce the interpretability of attention and feature maps. This challenge is especially evident in medical imaging, where domain shifts can degrade both performance and transparency. In this paper, we introduce Randomized-MLP (RMLP) regularization, a contrastive learning-based method that encourages more semantically aligned representations. We apply RMLP when fine-tuning DINOv2 to both medical and natural image modalities, showing that it improves or maintains downstream performance while producing more interpretable attention maps. We also provide a mathematical analysis of RMLPs, offering insights into its role in enhancing ViT-based models and advancing our understanding of contrastive learning in this context.

Building deep learning models that can rapidly segment whole slide images (WSIs) using only a handful of training samples remains an open challenge in computational pathology. The difficulty lies in the histological images themselves: many morphological structures within a slide are closely related and very similar in appearance, making it difficult to distinguish between them. However, a skilled pathologist can quickly identify the relevant phenotypes. Through years of training, they have learned to organize visual features into a hierarchical taxonomy (e.g., identifying carcinoma versus healthy tissue, or distinguishing regions within a tumor as cancer cells, the microenvironment, …). Thus, each region is associated with multiple labels representing different tissue types. Pathologists typically deal with this by analyzing the specimen at multiple scales and comparing visual features between different magnifications. Inspired by this multi-scale diagnostic workflow, we introduce the Navigator, a vision model that navigates through WSIs like a domain expert: it searches for the region of interest at a low scale, zooms in gradually, and localizes ever finer microanatomical classes. As a result, the Navigator can detect coarse-grained patterns at lower resolution and fine-grained features at higher resolution. In addition, to deal with sparsely annotated samples, we train the Navigator with a novel semi-supervised framework called S5CL v2. The proposed model improves the F1 score by up to 8% on various datasets including our challenging new TCGA-COAD-30CLS and Erlangen cohorts.

Training multimodal foundation models is challenging due to the limited availability of multimodal datasets. While many public datasets pair images with text, few combine images with audio or text with audio. Even rarer are datasets that align all three modalities at once. Critical domains such as healthcare, infrastructure, or transportation are particularly affected by missing modalities. This makes it difficult to integrate all modalities into a large pre-trained neural network that can be used out-of-the-box or fine-tuned for different downstream tasks. We introduce LoReTTa ($\textbf{L}$inking m$\textbf{O}$dalities with a t$\textbf{R}$ansitive and commutativ$\textbf{E}$ pre-$\textbf{T}$raining s$\textbf{T}$r$\textbf{A}$tegy) to address this understudied problem. Our self-supervised framework unifies causal modeling and masked modeling with the rules of commutativity and transitivity. This allows us to transition within and between modalities. As a result, our pre-trained models are better at exploring the true underlying joint probability distribution. Given a dataset containing only the disjoint combinations $(A, B)$ and $(B, C)$, LoReTTa can model the relation $A \leftrightarrow C$ with $A \leftrightarrow B \leftrightarrow C$. In particular, we show that a transformer pre-trained with LoReTTa can handle any mixture of modalities at inference time, including the never-seen pair $(A, C)$ and the triplet $(A, B, C)$. We extensively evaluate our approach on a synthetic, medical, and reinforcement learning dataset. Across different domains, our universal multimodal transformer consistently outperforms strong baselines such as GPT, BERT, and CLIP on tasks involving the missing modality tuple.

Segmenting the fine structure of the mouse brain on magnetic resonance (MR) images is critical for delineating morphological regions, analyzing brain function, and understanding their relationships. Compared to a single MRI modality, multimodal MRI data provide complementary tissue features that can be exploited by deep learning models, resulting in better segmentation results. However, multimodal mouse brain MRI data is often lacking, making automatic segmentation of mouse brain fine structure a very challenging task. To address this issue, it is necessary to fuse multimodal MRI data to produce distinguished contrasts in different brain structures. Hence, we propose a novel disentangled and contrastive GAN-based framework, named MouseGAN++, to synthesize multiple MR modalities from single ones in a structure-preserving manner, thus improving the segmentation performance by imputing missing modalities and multi-modality fusion. Our results demonstrate that the translation performance of our method outperforms the state-of-the-art methods. Using the subsequently learned modality-invariant information as well as the modality-translated images, MouseGAN++ can segment fine brain structures with averaged dice coefficients of 90.0% (T2w) and 87.9% (T1w), respectively, achieving around +10% performance improvement compared to the state-of-the-art algorithms. Our results demonstrate that MouseGAN++, as a simultaneous image synthesis and segmentation method, can be used to fuse cross-modality information in an unpaired manner and yield more robust performance in the absence of multimodal data. We release our method as a mouse brain structural segmentation tool for free academic usage at https://github.com/yu02019.

In recently published clinical trial results, hypoxia-modified therapies have shown to provide more positive outcomes to cancer patients, compared with standard cancer treatments. The development and validation of these hypoxia-modified therapies depend on an effective way of measuring tumor hypoxia, but a standardized measurement is currently unavailable in clinical practice. Different types of manual measurements have been proposed in clinical research, but in this paper we focus on a recently published approach that quantifies the number and proportion of hypoxic regions using high resolution (immuno-)fluorescence (IF) and hematoxylin and eosin (HE) stained images of a histological specimen of a tumor. We introduce new machine learning-based methodologies to automate this measurement, where the main challenge is the fact that the clinical annotations available for training the proposed methodologies consist of the total number of normoxic, chronically hypoxic, and acutely hypoxic regions without any indication of their location in the image. Therefore, this represents a weakly-supervised structured output classification problem, where training is based on a high-order loss function formed by the norm of the difference between the manual and estimated annotations mentioned above. We propose four methodologies to solve this problem: 1) a naive method that uses a majority classifier applied on the nodes of a fixed grid placed over the input images; 2) a baseline method based on a structured output learning formulation that relies on a fixed grid placed over the input images; 3) an extension to this baseline based on a latent structured output learning formulation that uses a graph that is flexible in terms of the amount and positions of nodes; and 4) a pixel-wise labeling based on a fully-convolutional neural network. Using a data set of 89 weakly annotated pairs of IF and HE images from eight tumors, we show that the quantitative results of methods (3) and (4) above are equally competitive and superior to the naive (1) and baseline (2) methods. All proposed methodologies show high correlation values with respect to the clinical annotations.

Staining and scanning of tissue samples for microscopic examination is fraught with undesirable color variations arising from differences in raw materials and manufacturing techniques of stain vendors, staining protocols of labs, and color responses of digital scanners. When comparing tissue samples, color normalization and stain separation of the tissue images can be helpful for both pathologists and software. Techniques that are used for natural images fail to utilize structural properties of stained tissue samples and produce undesirable color distortions. The stain concentration cannot be negative. Tissue samples are stained with only a few stains and most tissue regions are characterized by at most one effective stain. We model these physical phenomena that define the tissue structure by first decomposing images in an unsupervised manner into stain density maps that are sparse and non-negative. For a given image, we combine its stain density maps with stain color basis of a pathologist-preferred target image, thus altering only its color while preserving its structure described by the maps. Stain density correlation with ground truth and preference by pathologists were higher for images normalized using our method when compared to other alternatives. We also propose a computationally faster extension of this technique for large whole-slide images that selects an appropriate patch sample instead of using the entire image to compute the stain color basis.

We introduce two new structured output models that use a latent graph, which is flexible in terms of the number of nodes and structure, where the training process minimises a high-order loss function using a weakly annotated training set. These models are developed in the context of microscopy imaging of malignant tumours, where the estimation of the number and proportion of classes of microcirculatory supply units (MCSU) is important in the assessment of the efficacy of common cancer treatments (an MCSU is a region of the tumour tissue supplied by a microvessel). The proposed methodologies take as input multimodal microscopy images of a tumour, and estimate the number and proportion of MCSU classes. This estimation is facilitated by the use of an underlying latent graph (not present in the manual annotations), where each MCSU is represented by a node in this graph, labelled with the MCSU class and image location. The training process uses the manual weak annotations available, consisting of the number of MCSU classes per training image, where the training objective is the minimisation of a high-order loss function based on the norm of the error between the manual and estimated annotations. One of the models proposed is based on a new flexible latent structure support vector machine (FLSSVM) and the other is based on a deep convolutional neural network (DCNN) model. Using a dataset of 89 weakly annotated pairs of multimodal images from eight tumours, we show that the quantitative results from DCNN are superior, but the qualitative results from FLSSVM are better and both display high correlation values regarding the number and proportion of MCSU classes compared to the manual annotations.

The efficacy of cancer treatments (e.g., radiotherapy, chemotherapy, etc.) has been observed to critically depend on the proportion of hypoxic regions (i.e., a region deprived of adequate oxygen supply) in tumor tissue, so it is important to estimate this proportion from histological samples. Medical imaging data can be used to classify tumor tissue regions into necrotic or vital and then the vital tissue into normoxia (i.e., a region receiving a normal level of oxygen), chronic or acute hypoxia. Currently, this classification is a lengthy manual process performed using (immuno-)fluorescence (IF) and hematoxylin and eosin (HE) stained images of a histological specimen, which requires an expertise that is not widespread in clinical practice. In this paper, we propose a fully automated way to detect and classify tumor tissue regions into necrosis, normoxia, chronic hypoxia and acute hypoxia using IF and HE images from the same histological specimen. Instead of relying on any single classification methodology, we propose a principled combination of the following current state-of-the-art classifiers in the field: Adaboost, support vector machine, random forest and convolutional neural networks. Results show that on average we can successfully detect and classify more than 87% of the tumor tissue regions correctly. This automated system for estimating the proportion of chronic and acute hypoxia could provide clinicians with valuable information on assessing the efficacy of cancer treatments.