Built to Last? Reproducibility and Reusability of Deep Learning Algorithms in Computational Pathology.

Recent progress in computational pathology has been driven by deep learning. While code and data availability are essential to reproduce findings from preceding publications, ensuring a deep learning model's reusability is more challenging. For that, the codebase should be well-documented and easy to integrate into existing workflows and models should be robust toward noise and generalizable toward data from different sources. Strikingly, only a few computational pathology algorithms have been reused by other researchers so far, let alone employed in a clinical setting. To assess the current state of reproducibility and reusability of computational pathology algorithms, we evaluated peer-reviewed articles available in PubMed, published between January 2019 and March 2021, in 5 use cases: stain normalization; tissue type segmentation; evaluation of cell-level features; genetic alteration prediction; and inference of grading, staging, and prognostic information. We compiled criteria for data and code availability and statistical result analysis and assessed them in 160 publications. We found that only one-quarter (41 of 160 publications) made code publicly available. Among these 41 studies, three-quarters (30 of 41) analyzed their results statistically, half of them (20 of 41) released their trained model weights, and approximately a third (16 of 41) used an independent cohort for evaluation. Our review is intended for both pathologists interested in deep learning and researchers applying algorithms to computational pathology challenges. We provide a detailed overview of publications with published code in the field, list reusable data handling tools, and provide criteria for reproducibility and reusability.

RedTell: an AI tool for interpretable analysis of red blood cell morphology.

Introduction: Hematologists analyze microscopic images of red blood cells to study their morphology and functionality, detect disorders and search for drugs. However, accurate analysis of a large number of red blood cells needs automated computational approaches that rely on annotated datasets, expensive computational resources, and computer science expertise. We introduce RedTell, an AI tool for the interpretable analysis of red blood cell morphology comprising four single-cell modules: segmentation, feature extraction, assistance in data annotation, and classification. Methods: Cell segmentation is performed by a trained Mask R-CNN working robustly on a wide range of datasets requiring no or minimum fine-tuning. Over 130 features that are regularly used in research are extracted for every detected red blood cell. If required, users can train task-specific, highly accurate decision tree-based classifiers to categorize cells, requiring a minimal number of annotations and providing interpretable feature importance. Results: We demonstrate RedTell's applicability and power in three case studies. In the first case study we analyze the difference of the extracted features between the cells coming from patients suffering from different diseases, in the second study we use RedTell to analyze the control samples and use the extracted features to classify cells into echinocytes, discocytes and stomatocytes and finally in the last use case we distinguish sickle cells in sickle cell disease patients. Discussion: We believe that RedTell can accelerate and standardize red blood cell research and help gain new insights into mechanisms, diagnosis, and treatment of red blood cell associated disorders.

Explainable AI identifies diagnostic cells of genetic AML subtypes.

Explainable AI is deemed essential for clinical applications as it allows rationalizing model predictions, helping to build trust between clinicians and automated decision support tools. We developed an inherently explainable AI model for the classification of acute myeloid leukemia subtypes from blood smears and found that high-attention cells identified by the model coincide with those labeled as diagnostically relevant by human experts. Based on over 80,000 single white blood cell images from digitized blood smears of 129 patients diagnosed with one of four WHO-defined genetic AML subtypes and 60 healthy controls, we trained SCEMILA, a single-cell based explainable multiple instance learning algorithm. SCEMILA could perfectly discriminate between AML patients and healthy controls and detected the APL subtype with an F1 score of 0.86±0.05 (mean±s.d., 5-fold cross-validation). Analyzing a novel multi-attention module, we confirmed that our algorithm focused with high concordance on the same AML-specific cells as human experts do. Applied to classify single cells, it is able to highlight subtype specific cells and deconvolve the composition of a patient's blood smear without the need of single-cell annotation of the training data. Our large AML genetic subtype dataset is publicly available, and an interactive online tool facilitates the exploration of data and predictions. SCEMILA enables a comparison of algorithmic and expert decision criteria and can present a detailed analysis of individual patient data, paving the way to deploy AI in the routine diagnostics for identifying hematopoietic neoplasms.

SHAPR predicts 3D cell shapes from 2D microscopic images.

Reconstruction of shapes and sizes of three-dimensional (3D) objects from two- dimensional (2D) information is an intensely studied subject in computer vision. We here consider the level of single cells and nuclei and present a neural network-based SHApe PRediction autoencoder. For proof-of-concept, SHAPR reconstructs 3D shapes of red blood cells from single view 2D confocal microscopy images more accurately than naïve stereological models and significantly increases the feature-based prediction of red blood cell types from F1 = 79% to F1 = 87.4%. Applied to 2D images containing spheroidal aggregates of densely grown human induced pluripotent stem cells, we find that SHAPR learns fundamental shape properties of cell nuclei and allows for prediction-based morphometry. Reducing imaging time and data storage, SHAPR will help to optimize and up-scale image-based high-throughput applications for biomedicine.

Donor Age and Red Cell Age Contribute to the Variance in Lorrca Indices in Healthy Donors for Next Generation Ektacytometry: A Pilot Study.

The ability of red blood cells (RBCs) to transport gases, their lifespan as well as their rheological properties invariably depend on the deformability, hydration, and membrane stability of these cells, which can be measured by Laser optical rotational red cell analyser (Lorrca® Maxsis, RR Mechatronics). The osmoscan mode of Lorrca is currently used in diagnosis of rare anemias in clinical laboratories. However, a broad range of normal values for healthy subjects reduces the sensitivity of this method for diagnosis of mild disease phenotype. In this pilot study, we explored the impact of age and gender of 45 healthy donors, as well as RBC age on the Lorrca indices. Whereas gender did not affect the Lorrca indices in our study, the age donors had a profound effect on the O_hyper parameter. To study the impact of RBC age on the osmoscan parameters, we have isolated low (L)-, medium (M)-, or high (H)- density fractions enriched with young, mature, and senescent RBCs, respectively, and evaluated the influence of RBC age-related properties, such as density, morphology, and redox state, on the osmoscan indices. As before, O_hyper was the most sensitive parameter, dropping markedly with an increase in RBC density and age. Senescence was associated with a decrease in deformability (EI_max) and tolerability to low and high osmolatites (Area). L-fraction was enriched with reticulocytes and cells with high projected area and EMA staining, but also contained a small number of cells small in projected area and most likely, terminally senescent. L-fraction was on average slightly less deformable than mature cells. The cells from the L-fraction produced more oxidants and NO than all other fractions. However, RBCs from the L-fraction contained maximal levels of reduced thiols compared to other fractions. Our study suggests that reference values for O_hyper should be age-stratified, and, most probably, corrected for the average RBC age. Further multi-center study is required to validate these suggestions before implementing them into clinical practice.