Jonas Ammeling

Marc Aubreville, Nikolas Stathonikos, Taryn A. Donovan et al.

Recognition of mitotic figures in histologic tumor specimens is highly relevant to patient outcome assessment. This task is challenging for algorithms and human experts alike, with deterioration of algorithmic performance under shifts in image representations. Considerable covariate shifts occur when assessment is performed on different tumor types, images are acquired using different digitization devices, or specimens are produced in different laboratories. This observation motivated the inception of the 2022 challenge on MItosis Domain Generalization (MIDOG 2022). The challenge provided annotated histologic tumor images from six different domains and evaluated the algorithmic approaches for mitotic figure detection provided by nine challenge participants on ten independent domains. Ground truth for mitotic figure detection was established in two ways: a three-expert consensus and an independent, immunohistochemistry-assisted set of labels. This work represents an overview of the challenge tasks, the algorithmic strategies employed by the participants, and potential factors contributing to their success. With an $F_1$ score of 0.764 for the top-performing team, we summarize that domain generalization across various tumor domains is possible with today's deep learning-based recognition pipelines. However, we also found that domain characteristics not present in the training set (feline as new species, spindle cell shape as new morphology and a new scanner) led to small but significant decreases in performance. When assessed against the immunohistochemistry-assisted reference standard, all methods resulted in reduced recall scores, but with only minor changes in the order of participants in the ranking.

Sweta Banerjee, Timo Gosch, Sara Hester et al.

The annotation of large scale histopathology image datasets remains a major bottleneck in developing robust deep learning models for clinically relevant tasks, such as mitotic figure classification. Folder-based annotation workflows are usually slow, fatiguing, and difficult to scale. To address these challenges, we introduce SWipeable ANnotations (SWAN), an open-source, MIT-licensed web application that enables intuitive image patch classification using a swiping gesture. SWAN supports both desktop and mobile platforms, offers real-time metadata capture, and allows flexible mapping of swipe gestures to class labels. In a pilot study with four pathologists annotating 600 mitotic figure image patches, we compared SWAN against a traditional folder-sorting workflow. SWAN enabled rapid annotations with pairwise percent agreement ranging from 86.52% to 93.68% (Cohen's Kappa = 0.61-0.80), while for the folder-based method, the pairwise percent agreement ranged from 86.98% to 91.32% (Cohen's Kappa = 0.63-0.75) for the task of classifying atypical versus normal mitotic figures, demonstrating high consistency between annotators and comparable performance. Participants rated the tool as highly usable and appreciated the ability to annotate on mobile devices. These results suggest that SWAN can accelerate image annotation while maintaining annotation quality, offering a scalable and user-friendly alternative to conventional workflows.

Jonathan Ganz, Karoline Lipnik, Jonas Ammeling et al.

Nucleolar organizer regions (NORs) are parts of the DNA that are involved in RNA transcription. Due to the silver affinity of associated proteins, argyrophilic NORs (AgNORs) can be visualized using silver-based staining. The average number of AgNORs per nucleus has been shown to be a prognostic factor for predicting the outcome of many tumors. Since manual detection of AgNORs is laborious, automation is of high interest. We present a deep learning-based pipeline for automatically determining the AgNOR-score from histopathological sections. An additional annotation experiment was conducted with six pathologists to provide an independent performance evaluation of our approach. Across all raters and images, we found a mean squared error of 0.054 between the AgNOR- scores of the experts and those of the model, indicating that our approach offers performance comparable to humans.

Marc Aubreville, Jonathan Ganz, Jonas Ammeling et al.

Mitotic activity is key for the assessment of malignancy in many tumors. Moreover, it has been demonstrated that the proportion of abnormal mitosis to normal mitosis is of prognostic significance. Atypical mitotic figures (MF) can be identified morphologically as having segregation abnormalities of the chromatids. In this work, we perform, for the first time, automatic subtyping of mitotic figures into normal and atypical categories according to characteristic morphological appearances of the different phases of mitosis. Using the publicly available MIDOG21 and TUPAC16 breast cancer mitosis datasets, two experts blindly subtyped mitotic figures into five morphological categories. Further, we set up a state-of-the-art object detection pipeline extending the anchor-free FCOS approach with a gated hierarchical subclassification branch. Our labeling experiment indicated that subtyping of mitotic figures is a challenging task and prone to inter-rater disagreement, which we found in 24.89% of MF. Using the more diverse MIDOG21 dataset for training and TUPAC16 for testing, we reached a mean overall average precision score of 0.552, a ROC AUC score of 0.833 for atypical/normal MF and a mean class-averaged ROC-AUC score of 0.977 for discriminating the different phases of cells undergoing mitosis.

Marc Aubreville, Zhaoya Pan, Matti Sievert et al.

The surgical removal of head and neck tumors requires safe margins, which are usually confirmed intraoperatively by means of frozen sections. This method is, in itself, an oversampling procedure, which has a relatively low sensitivity compared to the definitive tissue analysis on paraffin-embedded sections. Confocal laser endomicroscopy (CLE) is an in-vivo imaging technique that has shown its potential in the live optical biopsy of tissue. An automated analysis of this notoriously difficult to interpret modality would help surgeons. However, the images of CLE show a wide variability of patterns, caused both by individual factors but also, and most strongly, by the anatomical structures of the imaged tissue, making it a challenging pattern recognition task. In this work, we evaluate four popular few shot learning (FSL) methods towards their capability of generalizing to unseen anatomical domains in CLE images. We evaluate this on images of sinunasal tumors (SNT) from five patients and on images of the vocal folds (VF) from 11 patients using a cross-validation scheme. The best respective approach reached a median accuracy of 79.6% on the rather homogeneous VF dataset, but only of 61.6% for the highly diverse SNT dataset. Our results indicate that FSL on CLE images is viable, but strongly affected by the number of patients, as well as the diversity of anatomical patterns.

Jonas Ammeling, Lars-Henning Schmidt, Jonathan Ganz et al.

Attention-based multiple instance learning (AMIL) algorithms have proven to be successful in utilizing gigapixel whole-slide images (WSIs) for a variety of different computational pathology tasks such as outcome prediction and cancer subtyping problems. We extended an AMIL approach to the task of survival prediction by utilizing the classical Cox partial likelihood as a loss function, converting the AMIL model into a nonlinear proportional hazards model. We applied the model to tissue microarray (TMA) slides of 330 lung cancer patients. The results show that AMIL approaches can handle very small amounts of tissue from a TMA and reach similar C-index performance compared to established survival prediction methods trained with highly discriminative clinical factors such as age, cancer grade, and cancer stage

Christof A. Bertram, Jonas Ammeling, Alexander Bartel et al.

Deep learning-based automated image analysis (DL-AIA) has been shown to outperform trained pathologists in tasks related to feature quantification. Related to these capacities the use of DL-AIA tools is currently extending from proof-of-principle studies to routine applications such as patient samples (diagnostic pathology), regulatory safety assessment (toxicologic pathology), and recurrent research tasks. To ensure that DL-AIA applications are safe and reliable, it is critical to conduct a thorough and objective generalization performance assessment (i.e., the ability of the algorithm to accurately predict patterns of interest) and possibly evaluate model robustness (i.e., the algorithm's capacity to maintain predictive accuracy on images from different sources). In this article, we review the practices for performance assessment in veterinary pathology publications by which two approaches were identified: 1) Exclusive visual performance control (i.e. eyeballing of algorithmic predictions) plus validation of the models application utilizing secondary performance indices, and 2) Statistical performance control (alongside the other methods), which requires a dataset creation and separation of an hold-out test set prior to model training. This article compares the strengths and weaknesses of statistical and visual performance control methods. Furthermore, we discuss relevant considerations for rigorous statistical performance evaluation including metric selection, test dataset image composition, ground truth label quality, resampling methods such as bootstrapping, statistical comparison of multiple models, and evaluation of model stability. It is our conclusion that visual and statistical evaluation have complementary strength and a combination of both provides the greatest insight into the DL model's performance and sources of error.

Emely Rosbach, Jonas Ammeling, Jonathan Ganz et al.

Artificial intelligence (AI)-driven decision support systems can improve diagnostic accuracy and efficiency in computational pathology. However, collaboration between human experts and AI may introduce cognitive biases such as automation and anchoring bias, where users adopt system predictions blindly or are disproportionately influenced by AI advice, even when inaccurate. These effects may be amplified under time pressure, common in routine pathology, or shaped by individual user characteristics. We conducted an online experiment in which pathology experts (n = 28) estimated tumor cell percentages: once independently and once with AI support. A subset of estimations in each condition was performed under time strain. Overall, AI assistance improved diagnostic performance but introduced a 7% automation bias rate, defined as accepted negative consultations where previously correct independent judgments were overturned by incorrect AI advice. While time pressure did not increase the frequency of automation bias, it appeared to intensify its severity, reflected in stronger performance declines associated with increased AI reliance under cognitive load. A linear mixed-effects model (LMM) simulating weighted averaging showed a statistically significant positive coefficient for AI advice, indicating moderate anchoring on system output. This effect increased under time pressure, suggesting anchoring bias becomes more pronounced when cognitive resources are limited. A second LMM assessing automation reliance, a proxy for automation and anchoring bias, showed that professional experience and self-efficacy were associated with lower dependence on AI, whereas higher confidence during AI-assisted decisions was tied to increased AI reliance. These findings highlight the dual nature of AI integration in clinical workflows: improving performance while introducing risks of bias-driven diagnostic errors.

Christof A. Bertram, Viktoria Weiss, Jonas Ammeling et al.

Supervised deep learning (DL) receives great interest for automated analysis of microscopic images with an increasing body of literature supporting its potential. The development and validation of those DL models relies heavily on the availability of high-quality, large-scale datasets. However, creating such datasets is a complex and resource-intensive process, often hindered by challenges such as time constraints, domain variability, and risks of bias in image collection and label creation. This review provides a comprehensive guide to the critical steps in dataset creation, including: 1) image acquisition, 2) selection of annotation software, and 3) annotation creation. In addition to ensuring a sufficiently large number of images, it is crucial to address sources of image variability (domain shifts) - such as those related to slide preparation and digitization - that could lead to algorithmic errors if not adequately represented in the training data. Key quality criteria for annotations are the three "C"s: correctness, completeness, and consistency. This review explores methods to enhance annotation quality through the use of advanced techniques that mitigate the limitations of single annotators. To support dataset creators, a standard operating procedure (SOP) is provided as supplemental material, outlining best practices for dataset development. Furthermore, the article underscores the importance of open datasets in driving innovation and enhancing reproducibility of DL research. By addressing the challenges and offering practical recommendations, this review aims to advance the creation of and availability to high-quality, large-scale datasets, ultimately contributing to the development of generalizable and robust DL models for pathology applications.

Jonas Ammeling, Moritz Hecker, Jonathan Ganz et al.

The volume-corrected mitotic index (M/V-Index) was shown to provide prognostic value in invasive breast carcinomas. However, despite its prognostic significance, it is not established as the standard method for assessing aggressive biological behaviour, due to the high additional workload associated with determining the epithelial proportion. In this work, we show that using a deep learning pipeline solely trained with an annotation-free, immunohistochemistry-based approach, provides accurate estimations of epithelial segmentation in canine breast carcinomas. We compare our automatic framework with the manually annotated M/V-Index in a study with three board-certified pathologists. Our results indicate that the deep learning-based pipeline shows expert-level performance, while providing time efficiency and reproducibility.

Sweta Banerjee, Viktoria Weiss, Taryn A. Donovan et al.

Atypical mitosis marks a deviation in the cell division process that has been shown be an independent prognostic marker for tumor malignancy. However, atypical mitosis classification remains challenging due to low prevalence, at times subtle morphological differences from normal mitotic figures, low inter-rater agreement among pathologists, and class imbalance in datasets. Building on the Atypical Mitosis dataset for Breast Cancer (AMi-Br), this study presents a comprehensive benchmark comparing deep learning approaches for automated atypical mitotic figure (AMF) classification, including end-to-end trained deep learning models, foundation models with linear probing, and foundation models fine-tuned with low-rank adaptation (LoRA). For rigorous evaluation, we further introduce two new held-out AMF datasets - AtNorM-Br, a dataset of mitotic figures from the TCGA breast cancer cohort, and AtNorM-MD, a multi-domain dataset of mitotic figures from a subset of the MIDOG++ training set. We found average balanced accuracy values of up to 0.8135, 0.7788, and 0.7723 on the in-domain AMi-Br and the out-of-domain AtNorm-Br and AtNorM-MD datasets, respectively. Our work shows that atypical mitotic figure classification, while being a challenging problem, can be effectively addressed through the use of recent advances in transfer learning and model fine-tuning techniques. We make all code and data used in this paper available in this github repository: https://github.com/DeepMicroscopy/AMi-Br_Benchmark.

Jingna Qiu, Nishanth Jain, Jonas Ammeling et al.

Recent advances in Vision-Language Models (VLMs) in histopathology, such as CONCH and QuiltNet, have demonstrated impressive zero-shot classification capabilities across various tasks. However, their general-purpose design may lead to suboptimal performance in specific downstream applications. While supervised fine-tuning methods address this issue, they require manually labeled samples for adaptation. This paper investigates annotation-free adaptation of VLMs through continued pretraining on domain- and task-relevant image-caption pairs extracted from existing databases. Our experiments on two VLMs, CONCH and QuiltNet, across three downstream tasks reveal that these pairs substantially enhance both zero-shot and few-shot performance. Notably, with larger training sizes, continued pretraining matches the performance of few-shot methods while eliminating manual labeling. Its effectiveness, task-agnostic design, and annotation-free workflow make it a promising pathway for adapting VLMs to new histopathology tasks. Code is available at https://github.com/DeepMicroscopy/Annotation-free-VLM-specialization.

Christof A. Bertram, Viktoria Weiss, Taryn A. Donovan et al.

Assessment of the density of mitotic figures (MFs) in histologic tumor sections is an important prognostic marker for many tumor types, including breast cancer. Recently, it has been reported in multiple works that the quantity of MFs with an atypical morphology (atypical MFs, AMFs) might be an independent prognostic criterion for breast cancer. AMFs are an indicator of mutations in the genes regulating the cell cycle and can lead to aberrant chromosome constitution (aneuploidy) of the tumor cells. To facilitate further research on this topic using pattern recognition, we present the first ever publicly available dataset of atypical and normal MFs (AMi-Br). For this, we utilized two of the most popular MF datasets (MIDOG 2021 and TUPAC) and subclassified all MFs using a three expert majority vote. Our final dataset consists of 3,720 MFs, split into 832 AMFs (22.4%) and 2,888 normal MFs (77.6%) across all 223 tumor cases in the combined set. We provide baseline classification experiments to investigate the consistency of the dataset, using a Monte Carlo cross-validation and different strategies to combat class imbalance. We found an averaged balanced accuracy of up to 0.806 when using a patch-level data set split, and up to 0.713 when using a patient-level split.

Marc Aubreville, Jonathan Ganz, Jonas Ammeling et al.

The QUILT-1M dataset is the first openly available dataset containing images harvested from various online sources. While it provides a huge data variety, the image quality and composition is highly heterogeneous, impacting its utility for text-conditional image synthesis. We propose an automatic pipeline that provides predictions of the most common impurities within the images, e.g., visibility of narrators, desktop environment and pathology software, or text within the image. Additionally, we propose to use semantic alignment filtering of the image-text pairs. Our findings demonstrate that by rigorously filtering the dataset, there is a substantial enhancement of image fidelity in text-to-image tasks.

Frauke Wilm, Jonas Ammeling, Mathias Öttl et al.

The U-net architecture has significantly impacted deep learning-based segmentation of medical images. Through the integration of long-range skip connections, it facilitated the preservation of high-resolution features. Out-of-distribution data can, however, substantially impede the performance of neural networks. Previous works showed that the trained network layers differ in their susceptibility to this domain shift, e.g., shallow layers are more affected than deeper layers. In this work, we investigate the implications of this observation of layer sensitivity to domain shifts of U-net-style segmentation networks. By copying features of shallow layers to corresponding decoder blocks, these bear the risk of re-introducing domain-specific information. We used a synthetic dataset to model different levels of data distribution shifts and evaluated the impact on downstream segmentation performance. We quantified the inherent domain susceptibility of each network layer, using the Hellinger distance. These experiments confirmed the higher domain susceptibility of earlier network layers. When gradually removing skip connections, a decrease in domain susceptibility of deeper layers could be observed. For downstream segmentation performance, the original U-net outperformed the variant without any skip connections. The best performance, however, was achieved when removing the uppermost skip connection - not only in the presence of domain shifts but also for in-domain test data. We validated our results on three clinical datasets - two histopathology datasets and one magnetic resonance dataset - with performance increases of up to 10% in-domain and 13% cross-domain when removing the uppermost skip connection.

Jonas Ammeling, Jonathan Ganz, Emely Rosbach et al.

The performance of deep learning models is known to scale with data quantity and diversity. In pathology, as in many other medical imaging domains, the availability of labeled images for a specific task is often limited. Self-supervised learning techniques have enabled the use of vast amounts of unlabeled data to train large-scale neural networks, i.e., foundation models, that can address the limited data problem by providing semantically rich feature vectors that can generalize well to new tasks with minimal training effort increasing model performance and robustness. In this work, we investigate the use of foundation models for mitotic figure classification. The mitotic count, which can be derived from this classification task, is an independent prognostic marker for specific tumors and part of certain tumor grading systems. In particular, we investigate the data scaling laws on multiple current foundation models and evaluate their robustness to unseen tumor domains. Next to the commonly used linear probing paradigm, we also adapt the models using low-rank adaptation (LoRA) of their attention mechanisms. We compare all models against end-to-end-trained baselines, both CNNs and Vision Transformers. Our results demonstrate that LoRA-adapted foundation models provide superior performance to those adapted with standard linear probing, reaching performance levels close to 100% data availability with only 10% of training data. Furthermore, LoRA-adaptation of the most recent foundation models almost closes the out-of-domain performance gap when evaluated on unseen tumor domains. However, full fine-tuning of traditional architectures still yields competitive performance.

Jonathan Ganz, Jonas Ammeling, Emely Rosbach et al.

Foundation models (FMs), i.e., models trained on a vast amount of typically unlabeled data, have become popular and available recently for the domain of histopathology. The key idea is to extract semantically rich vectors from any input patch, allowing for the use of simple subsequent classification networks potentially reducing the required amounts of labeled data, and increasing domain robustness. In this work, we investigate to which degree this also holds for mitotic figure classification. Utilizing two popular public mitotic figure datasets, we compared linear probing of five publicly available FMs against models trained on ImageNet and a simple ResNet50 end-to-end-trained baseline. We found that the end-to-end-trained baseline outperformed all FM-based classifiers, regardless of the amount of data provided. Additionally, we did not observe the FM-based classifiers to be more robust against domain shifts, rendering both of the above assumptions incorrect.

Jonathan Ganz, Christian Marzahl, Jonas Ammeling et al.

The count of mitotic figures (MFs) observed in hematoxylin and eosin (H&E)-stained slides is an important prognostic marker as it is a measure for tumor cell proliferation. However, the identification of MFs has a known low inter-rater agreement. Deep learning algorithms can standardize this task, but they require large amounts of annotated data for training and validation. Furthermore, label noise introduced during the annotation process may impede the algorithm's performance. Unlike H&E, the mitosis-specific antibody phospho-histone H3 (PHH3) specifically highlights MFs. Counting MFs on slides stained against PHH3 leads to higher agreement among raters and has therefore recently been used as a ground truth for the annotation of MFs in H&E. However, as PHH3 facilitates the recognition of cells indistinguishable from H&E stain alone, the use of this ground truth could potentially introduce noise into the H&E-related dataset, impacting model performance. This study analyzes the impact of PHH3-assisted MF annotation on inter-rater reliability and object level agreement through an extensive multi-rater experiment. We found that the annotators' object-level agreement increased when using PHH3-assisted labeling. Subsequently, MF detectors were evaluated on the resulting datasets to investigate the influence of PHH3-assisted labeling on the models' performance. Additionally, a novel dual-stain MF detector was developed to investigate the interpretation-shift of PHH3-assisted labels used in H&E, which clearly outperformed single-stain detectors. However, the PHH3-assisted labels did not have a positive effect on solely H&E-based models. The high performance of our dual-input detector reveals an information mismatch between the H&E and PHH3-stained images as the cause of this effect.

Jonathan Ganz, Jonas Ammeling, Samir Jabari et al.

In numerous studies, deep learning algorithms have proven their potential for the analysis of histopathology images, for example, for revealing the subtypes of tumors or the primary origin of metastases. These models require large datasets for training, which must be anonymized to prevent possible patient identity leaks. This study demonstrates that even relatively simple deep learning algorithms can re-identify patients in large histopathology datasets with substantial accuracy. We evaluated our algorithms on two TCIA datasets including lung squamous cell carcinoma (LSCC) and lung adenocarcinoma (LUAD). We also demonstrate the algorithm's performance on an in-house dataset of meningioma tissue. We predicted the source patient of a slide with F1 scores of 50.16 % and 52.30 % on the LSCC and LUAD datasets, respectively, and with 62.31 % on our meningioma dataset. Based on our findings, we formulated a risk assessment scheme to estimate the risk to the patient's privacy prior to publication.