Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning. Our results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.

Multiple Instance Learning (MIL) is a cornerstone approach in computational pathology for distilling embeddings from gigapixel tissue images into patient-level representations to predict clinical outcomes. However, MIL is frequently challenged by the constraints of working with small, weakly-supervised clinical datasets. Unlike fields such as natural language processing and computer vision, which effectively use transfer learning to improve model quality in data-scarce environments, the transferability of MIL models remains largely unexplored. We conduct the first comprehensive investigation into transfer learning capabilities of pretrained MIL models, evaluating 11 MIL models across 19 pretraining tasks spanning tissue subtyping, cancer grading, and molecular subtype prediction. We observe a substantial performance boost with finetuning pretrained models over training from randomly initialized weights, even with domain differences between pretraining and target tasks. Pretraining on pan-cancer datasets enables consistent generalization across organs and task types compared to single-disease pretraining. Remarkably, this pan-cancer pretraining leads to better transfer than that of a state-of-the-art slide-level foundation model, while using only 6.5\% of the training data. These findings indicate that MIL architectures exhibit robust adaptability, offering insights into the benefits of leveraging pretrained models to enhance performance in computational pathology.

Self-supervised learning (SSL) has been successful in building patch embeddings of small histology images (e.g.,$224\times 224$pixels), but scaling these models to learn slide embeddings from the entirety of giga-pixel whole-slide images (WSIs) remains challenging. Here, we leverage complementary information from gene expression profiles to guide slide representation learning using multi-modal pre-training. Expression profiles constitute highly detailed molecular descriptions of a tissue that we hypothesize offer a strong task-agnostic training signal for learning slide embeddings. Our slide and expression (S+E) pre-training strategy, called Tangle, employs modality-specific encoders, the outputs of which are aligned via contrastive learning. Tangle was pre-trained on samples from three different organs: liver (n=6,597 S+E pairs), breast (n=1,020), and lung (n=1,012) from two different species (Homo sapiens and Rattus norvegicus). Across three independent test datasets consisting of 1,265 breast WSIs, 1,946 lung WSIs, and 4,584 liver WSIs, Tangle shows significantly better few-shot performance compared to supervised and SSL baselines. When assessed using prototype-based classification and slide retrieval, Tangle also shows a substantial performance improvement over all baselines. Code available at https://github.com/mahmoodlab/TANGLE.

Integrating whole-slide images (WSIs) and bulk tran-scriptomics for predicting patient survival can improve our understanding of patient prognosis. However, this multi-modal task is particularly challenging due to the different nature of these data: WSIs represent a very high-dimensional spatial description of a tumor, while bulk tran-scriptomics represent a global description of gene expression levels within that tumor. In this context, our work aims to address two key challenges: (1) how can we tokenize transcriptomics in a semantically meaningful and interpretable way?, and (2) how can we capture dense multi-modal interactions between these two modalities? Here, we propose to learn biological pathway tokens from transcriptomics that can encode specific cellular functions. Together with histology patch tokens that encode the slide morphology, we argue that they form appropriate reasoning units for interpretability. We fuse both modalities using a memoryefficient multimodal Transformer that can model interactions between pathway and histology patch tokens. Our model, Survpath, achieves state-of-the-art performance when evaluated against unimodal and multimodal baselines on five datasets from The Cancer Genome Atlas. Our interpretability framework identifies key multimodal prognostic factors, and, as such, can provide valuable insights into the interaction between genotype and phenotype. Code available at https://github.com/mahmoodlab/SurvPath.

Representation learning of pathology whole-slide images (WSIs) has been has primarily relied on weak supervision with Multiple Instance Learning (MIL). However, the slide representations resulting from this approach are highly tailored to specific clinical tasks, which limits their expressivity and generalization, particularly in scenarios with limited data. Instead, we hypothesize that morphological redundancy in tissue can be leveraged to build a task-agnostic slide representation in an unsupervised fashion. To this end, we introduce Panther, a prototype-based approach rooted in the Gaussian mixture model that summarizes the set of WSI patches into a much smaller set of morphological prototypes. Specifically, each patch is assumed to have been generated from a mixture distri-bution, where each mixture component represents a morphological exemplar. Utilizing the estimated mixture parameters, we then construct a compact slide representation that can be readily used for a wide range of downstream tasks. By performing an extensive evaluation of Panther on subtyping and survival tasks using 13 datasets, we show that 1) Panther outperforms or is on par with super-vised MIL baselines and 2) the analysis of morphological prototypes brings new qualitative and quantitative in-sights into model interpretability. The code is available at https://github.com/mahmoodlab/Panther.

Multimodal survival methods combining gigapixel histology whole-slide images (WSIs) and transcriptomic profiles are particularly promising for patient prognostication and stratification. Current approaches involve tokenizing the WSIs into smaller patches ($>10^4$ patches) and transcriptomics into gene groups, which are then integrated using a Transformer for predicting outcomes. However, this process generates many tokens, which leads to high memory requirements for computing attention and complicates post-hoc interpretability analyses. Instead, we hypothesize that we can: (1) effectively summarize the morphological content of a WSI by condensing its constituting tokens using morphological prototypes, achieving more than $300\times$ compression; and (2) accurately characterize cellular functions by encoding the transcriptomic profile with biological pathway prototypes, all in an unsupervised fashion. The resulting multimodal tokens are then processed by a fusion network, either with a Transformer or an optimal transport cross-alignment, which now operates with a small and fixed number of tokens without approximations. Extensive evaluation on six cancer types shows that our framework outperforms state-of-the-art methods with much less computation while unlocking new interpretability analyses. The code is available at https://github.com/mahmoodlab/MMP.

Spatial transcriptomics enables interrogating the molecular composition of tissue with ever-increasing resolution and sensitivity. However, costs, rapidly evolving technology, and lack of standards have constrained computational methods in ST to narrow tasks and small cohorts. In addition, the underlying tissue morphology, as reflected by H&E-stained whole slide images (WSIs), encodes rich information often overlooked in ST studies. Here, we introduce HEST-1k, a collection of 1,229 spatial transcriptomic profiles, each linked to a WSI and extensive metadata. HEST-1k was assembled from 153 public and internal cohorts encompassing 26 organs, two species (Homo Sapiens and Mus Musculus), and 367 cancer samples from 25 cancer types. HEST-1k processing enabled the identification of 2.1 million expression-morphology pairs and over 76 million nuclei. To support its development, we additionally introduce the HEST-Library, a Python package designed to perform a range of actions with HEST samples. We test HEST-1k and Library on three use cases: (1) benchmarking foundation models for pathology (HEST-Benchmark), (2) biomarker exploration, and (3) multimodal representation learning. HEST-1k, HEST-Library, and HEST-Benchmark can be freely accessed at https://github.com/mahmoodlab/hest.

Contrastive visual language pretraining has emerged as a powerful method for either training new language-aware image encoders or augmenting existing pretrained models with zero-shot visual recognition capabilities. However, existing works typically train on large datasets of imagetext pairs and have been designed to perform downstream tasks involving only small to medium sized-images, neither of which are applicable to the emerging field of computational pathology where there are limited publicly available paired image-text datasets and each image can span up to 100,000 × 100,000 pixels. In this paper we present MI-Zero, a simple and intuitive framework for unleashing the zero-shot transfer capabilities of contrastively aligned image and text models on gigapixel histopathology whole slide images, enabling multiple downstream diagnostic tasks to be carried out by pretrained encoders without requiring any additional labels. MI-Zero reformulates zero-shot transfer under the framework of multiple instance learning to overcome the computational challenge of inference on extremely large images. We used over 550k pathology reports and other available in-domain text corpora to pretrain our text encoder. By effectively leveraging strong pretrained encoders, our best model pretrained on over 33k histopathology image-caption pairs achieves an average median zero-shot accuracy of 70.2% across three different real-world cancer subtyping tasks. Our code is available at: https://github.com/mahmoodlab/MI-Zero.

The recent explosion of interest in multimodal applications has resulted in a wide selection of datasets and methods for representing and integrating information from different modalities. Despite these empirical advances, there remain fundamental research questions: How can we quantify the interactions that are necessary to solve a multimodal task? Subsequently, what are the most suitable multimodal models to capture these interactions? To answer these questions, we propose an information-theoretic approach to quantify the degree of redundancy, uniqueness, and synergy relating input modalities with an output task. We term these three measures as the PID statistics of a multimodal distribution (or PID for short), and introduce two new estimators for these PID statistics that scale to high-dimensional distributions. To validate PID estimation, we conduct extensive experiments on both synthetic datasets where the PID is known and on large-scale multimodal benchmarks where PID estimations are compared with human annotations. Finally, we demonstrate their usefulness in (1) quantifying interactions within multimodal datasets, (2) quantifying interactions captured by multimodal models, (3) principled approaches for model selection, and (4) three real-world case studies engaging with domain experts in pathology, mood prediction, and robotic perception where our framework helps to recommend strong multimodal models for each application.

Vision Transformers (ViTs) and their multi-scale and hierarchical variations have been successful at capturing image representations but their use has been generally studied for low-resolution images (e.g. 256 × 256, 384 × 384). For gigapixel whole-slide imaging (WSI) in computational pathology, WSIs can be as large as 150000 × 150000 pixels at 20 × magnification and exhibit a hierarchical structure of visual tokens across varying resolutions: from 16 × 16 images capturing individual cells, to 4096 × 4096 images characterizing interactions within the tissue microenvironment. We introduce a new ViT architecture called the Hierarchical Image Pyramid Transformer (HIPT), which leverages the natural hierarchical structure inherent in WSIs using two levels of self-supervised learning to learn high-resolution image representations. HIPT is pretrained across 33 cancer types using 10,678 gigapixel WSIs, 408,218 4096 × 4096 images, and 104M 256 × 256 images. We benchmark HIPT representations on 9 slide-level tasks, and demonstrate that: 1) HIPT with hierarchical pretraining outperforms current state-of-the-art methods for cancer subtyping and survival prediction, 2) self-supervised ViTs are able to model important inductive biases about the hierarchical structure of phenotypes in the tumor microenvironment.

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.

Cancer diagnosis, prognosis, mymargin and therapeutic response predictions are based on morphological information from histology slides and molecular profiles from genomic data. However, most deep learning-based objective outcome prediction and grading paradigms are based on histology or genomics alone and do not make use of the complementary information in an intuitive manner. In this work, we propose Pathomic Fusion, an interpretable strategy for end-to-end multimodal fusion of histology image and genomic (mutations, CNV, RNA-Seq) features for survival outcome prediction. Our approach models pairwise feature interactions across modalities by taking the Kronecker product of unimodal feature representations, and controls the expressiveness of each representation via a gating-based attention mechanism. Following supervised learning, we are able to interpret and saliently localize features across each modality, and understand how feature importance shifts when conditioning on multimodal input. We validate our approach using glioma and clear cell renal cell carcinoma datasets from the Cancer Genome Atlas (TCGA), which contains paired whole-slide image, genotype, and transcriptome data with ground truth survival and histologic grade labels. In a 15-fold cross-validation, our results demonstrate that the proposed multimodal fusion paradigm improves prognostic determinations from ground truth grading and molecular subtyping, as well as unimodal deep networks trained on histology and genomic data alone. The proposed method establishes insight and theory on how to train deep networks on multimodal biomedical data in an intuitive manner, which will be useful for other problems in medicine that seek to combine heterogeneous data streams for understanding diseases and predicting response and resistance to treatment. Code and trained models are made available at: https://github.com/mahmoodlab/PathomicFusion.

Survival outcome prediction is a challenging weakly-supervised and ordinal regression task in computational pathology that involves modeling complex interactions within the tumor microenvironment in gigapixel whole slide images (WSIs). Despite recent progress in formulating WSIs as bags for multiple instance learning (MIL), representation learning of entire WSIs remains an open and challenging problem, especially in overcoming: 1) the computational complexity of feature aggregation in large bags, and 2) the data heterogeneity gap in incorporating biological priors such as genomic measurements. In this work, we present a Multimodal Co-Attention Transformer (MCAT) framework that learns an interpretable, dense co-attention mapping between WSIs and genomic features formulated in an embedding space. Inspired by approaches in Visual Question Answering (VQA) that can attribute how word embed-dings attend to salient objects in an image when answering a question, MCAT learns how histology patches attend to genes when predicting patient survival. In addition to visualizing multimodal interactions, our co-attention trans-formation also reduces the space complexity of WSI bags, which enables the adaptation of Transformer layers as a general encoder backbone in MIL. We apply our proposed method on five different cancer datasets (4,730 WSIs, 67 million patches). Our experimental results demonstrate that the proposed method consistently achieves superior performance compared to the state-of-the-art methods.

Although wireless capsule endoscopy is the preferred modality for diagnosis and assessment of small bowel diseases, the poor camera resolution is a substantial limitation for both subjective and automated diagnostics. Enhanced-resolution endoscopy has shown to improve adenoma detection rate for conventional endoscopy and is likely to do the same for capsule endoscopy. In this work, we propose and quantitatively validate a novel framework to learn a mapping from low-to-high-resolution endoscopic images. We combine conditional adversarial networks with a spatial attention block to improve the resolution by up to factors of 8× , 10× , 12× , respectively. Quantitative and qualitative studies demonstrate the superiority of EndoL2H over state-of-the-art deep super-resolution methods Deep Back-Projection Networks (DBPN), Deep Residual Channel Attention Networks (RCAN) and Super Resolution Generative Adversarial Network (SRGAN). Mean Opinion Score (MOS) tests were performed by 30 gastroenterologists qualitatively assess and confirm the clinical relevance of the approach. EndoL2H is generally applicable to any endoscopic capsule system and has the potential to improve diagnosis and better harness computational approaches for polyp detection and characterization. Our code and trained models are available at https://github.com/CapsuleEndoscope/EndoL2H.

We present a deep learning framework for wide-field, content-aware estimation of absorption and scattering coefficients of tissues, called Generative Adversarial Network Prediction of Optical Properties (GANPOP). Spatial frequency domain imaging is used to obtain ground-truth optical properties at 660 nm from in vivo human hands and feet, freshly resected human esophagectomy samples, and homogeneous tissue phantoms. Images of objects with either flat-field or structured illumination are paired with registered optical property maps and are used to train conditional generative adversarial networks that estimate optical properties from a single input image. We benchmark this approach by comparing GANPOP to a single-snapshot optical property (SSOP) technique, using a normalized mean absolute error (NMAE) metric. In human gastrointestinal specimens, GANPOP with a single structured-light input image estimates the reduced scattering and absorption coefficients with 60% higher accuracy than SSOP while GANPOP with a single flat-field illumination image achieves similar accuracy to SSOP. When applied to both in vivo and ex vivo swine tissues, a GANPOP model trained solely on structured-illumination images of human specimens and phantoms estimates optical properties with approximately 46% improvement over SSOP, indicating adaptability to new, unseen tissue types. Given a training set that appropriately spans the target domain, GANPOP has the potential to enable rapid and accurate wide-field measurements of optical properties.

Nuclei mymargin segmentation is a fundamental task for various computational pathology applications including nuclei morphology analysis, cell type classification, and cancer grading. Deep learning has emerged as a powerful approach to segmenting nuclei but the accuracy of convolutional neural networks (CNNs) depends on the volume and the quality of labeled histopathology data for training. In particular, conventional CNN-based approaches lack structured prediction capabilities, which are required to distinguish overlapping and clumped nuclei. Here, we present an approach to nuclei segmentation that overcomes these challenges by utilizing a conditional generative adversarial network (cGAN) trained with synthetic and real data. We generate a large dataset of H&E training images with perfect nuclei segmentation labels using an unpaired GAN framework. This synthetic data along with real histopathology data from six different organs are used to train a conditional GAN with spectral normalization and gradient penalty for nuclei segmentation. This adversarial regression framework enforces higher-order spacial-consistency when compared to conventional CNN models. We demonstrate that this nuclei segmentation approach generalizes across different organs, sites, patients and disease states, and outperforms conventional approaches, especially in isolating individual and overlapping nuclei.

Sparsity exploiting image reconstruction (SER) methods have been extensively used with total variation (TV) regularization for tomographic reconstructions. Local TV methods fail to preserve texture details and often create additional artifacts due to over-smoothing. Nonlocal TV (NLTV) methods have been proposed as a solution to this but they either lack continuous updates due to computational constraints or limit the locality to a small region. In this letter, we propose adaptive graph-based TV. The proposed method goes beyond spatial similarity between different regions of an image being reconstructed by establishing a connection between similar regions in the entire image regardless of spatial distance. As compared to NLTV, the proposed method is computationally efficient and involves updating the graph prior during every iteration making the connection between similar regions stronger. Moreover, it promotes sparsity in the wavelet and graph gradient domains. Since TV is a special case of graph TV, the proposed method can also be seen as a generalization of SER and TV methods.

To realize the full potential of deep learning for medical imaging, large annotated datasets are required for training. Such datasets are difficult to acquire due to privacy issues, lack of experts available for annotation, underrepresentation of rare conditions, and poor standardization. The lack of annotated data has been addressed in conventional vision applications using synthetic images refined via unsupervised adversarial training to look like real images. However, this approach is difficult to extend to general medical imaging because of the complex and diverse set of features found in real human tissues. We propose a novel framework that uses a reverse flow, where adversarial training is used to make real medical images more like synthetic images, and clinically-relevant features are preserved via self-regularization. These domain-adapted synthetic-like images can then be accurately interpreted by networks trained on large datasets of synthetic medical images. We implement this approach on the notoriously difficult task of depth-estimation from monocular endoscopy which has a variety of applications in colonoscopy, robotic surgery, and invasive endoscopic procedures. We train a depth estimator on a large data set of synthetic images generated using an accurate forward model of an endoscope and an anatomically-realistic colon. Our analysis demonstrates that the structural similarity of endoscopy depth estimation in a real pig colon predicted from a network trained solely on synthetic data improved by 78.7% by using reverse domain adaptation.

Two-Dimensional (2D) Discrete Fourier Transform (DFT) is a basic and computationally intensive algorithm, with a vast variety of applications. 2D images are, in general, non-periodic, but are assumed to be periodic while calculating their DFTs. This leads to cross-shaped artifacts in the frequency domain due to spectral leakage. These artifacts can have critical consequences if the DFTs are being used for further processing. In this paper we present a novel FPGA-based design to calculate high-throughput 2D DFTs with simultaneous edge artifact removal. Standard approaches for removing these artifacts using apodization functions or mirroring, either involve removing critical frequencies or a surge in computation by increasing image size. We use a periodic-plus-smooth decomposition based artifact removal algorithm optimized for FPGA implementation, while still achieving real-time (≥23 frames per second) performance for a 512×512 size image stream. Our optimization approach leads to a significant decrease in external memory utilization thereby avoiding memory conflicts and simplifies the design. We have tested our design on a PXIe based Xilinx Kintex 7 FPGA system communicating with a host PC which gives us the advantage to further expand the design for industrial applications.