Improving classification accuracy of fine-tuned CNN models: Impact of hyperparameter optimization.

The immense popularity of convolutional neural network (CNN) models has sparked a growing interest in optimizing their hyperparameters. Discovering the ideal values for hyperparameters to achieve optimal CNN training is a complex and time-consuming task, often requiring repetitive numerical experiments. As a result, significant attention is currently being devoted to developing methods aimed at tailoring hyperparameters for specific CNN models and classification tasks. While existing optimization methods often yield favorable image classification results, they do not provide guidance on which hyperparameters are worth optimizing, the appropriate value ranges for those hyperparameters, or whether it is reasonable to use a subset of training data for the optimization process. This work is focused on the optimization of hyperparameters during transfer learning, with the goal of investigating how different optimization methods and hyperparameter selections impact the performance of fine-tuned models. In our experiments, we assessed the importance of various hyperparameters and identified the ranges within which optimal CNN training can be achieved. Additionally, we compared four hyperparameter optimization methods-grid search, random search, Bayesian optimization, and the Asynchronous Successive Halving Algorithm (ASHA). We also explored the feasibility of fine-tuning hyperparameters using a subset of the training data. By optimizing the hyperparameters, we observed an improvement in CNN classification accuracy of up to 6%. Furthermore, we found that achieving a balance in class distribution within the subset of data used for parameter optimization is crucial in establishing the optimal set of hyperparameters for CNN training. The results we obtained demonstrate that hyperparameter optimization is highly dependent on the specific task and dataset at hand.

The Molecular Twin artificial-intelligence platform integrates multi-omic data to predict outcomes for pancreatic adenocarcinoma patients.

Contemporary analyses focused on a limited number of clinical and molecular biomarkers have been unable to accurately predict clinical outcomes in pancreatic ductal adenocarcinoma. Here we describe a precision medicine platform known as the Molecular Twin consisting of advanced machine-learning models and use it to analyze a dataset of 6,363 clinical and multi-omic molecular features from patients with resected pancreatic ductal adenocarcinoma to accurately predict disease survival (DS). We show that a full multi-omic model predicts DS with the highest accuracy and that plasma protein is the top single-omic predictor of DS. A parsimonious model learning only 589 multi-omic features demonstrated similar predictive performance as the full multi-omic model. Our platform enables discovery of parsimonious biomarker panels and performance assessment of outcome prediction models learning from resource-intensive panels. This approach has considerable potential to impact clinical care and democratize precision cancer medicine worldwide.

Selective Growth of van der Waals Heterostructures Enabled by Electron-Beam Irradiation.

Van der Waals heterostructures (vdWHSs) enable the fabrication of complex electronic devices based on two-dimensional (2D) materials. Ideally, these vdWHSs should be fabricated in a scalable and repeatable way and only in the specific areas of the substrate to lower the number of technological operations inducing defects and impurities. Here, we present a method of selective fabrication of vdWHSs via chemical vapor deposition by electron-beam (EB) irradiation. We distinguish two growth modes: positive (2D materials nucleate on the irradiated regions) on graphene and tungsten disulfide (WS2) substrates, and negative (2D materials do not nucleate on the irradiated regions) on the graphene substrate. The growth mode is controlled by limiting the air exposure of the irradiated substrate and the time between irradiation and growth. We conducted Raman mapping, Kelvin-probe force microscopy, X-ray photoelectron spectroscopy, and density-functional theory modeling studies to investigate the selective growth mechanism. We conclude that the selective growth is explained by the competition of three effects: EB-induced defects, adsorption of carbon species, and electrostatic interaction. The method here is a critical step toward the industry-scale fabrication of 2D-materials-based devices.

Multilayer outperforms single-layer slide scanning in AI-based classification of whole slide images with low-burden acid-fast mycobacteria (AFB).

Manual screening of Ziehl-Neelsen (ZN)-stained slides that are negative or contain rare acid-fast mycobacteria (AFB) is labor-intensive and requires repetitive refocusing to visualize AFB candidates under the microscope. Whole slide image (WSI) scanners have enabled implementation of AI to classify digital ZN-stained slides as AFB+ or AFB-. By default, these scanners acquire a single-layer WSI. However, some scanners can acquire a multilayer WSI with a z-stack and an extended focus image layer embedded. We developed a parameterized WSI classification pipeline to assess whether multilayer imaging improves ZN-stained slide classification accuracy. A CNN built into the pipeline classified tiles in each image layer to form an AFB probability score heatmap. Features extracted from the heatmap were then entered into a WSI classifier. 46 AFB+ and 88 AFB- single-layer WSIs were used for the classifier training. 15 AFB+ (with rare microorganisms) and 5 AFB- multilayer WSIs comprised the test set. Parameters in the pipeline included: (a) a WSI representation: z-stack of image layers, middle image layer (a single image layer equivalent) or an extended focus image layer, (b) 4 methods of aggregating AFB probability scores across the z-stack, (c) 3 classifiers, (d) 3 AFB probability thresholds, and (e) 9 feature vector types extracted from the aggregated AFB probability heatmaps. Balanced accuracy (BACC) was used to measure the pipeline performance for all parameter combinations. Analysis of Covariance (ANCOVA) was used to statistically evaluate the effect of each parameter on the BACC. After adjusting for other factors, a significant effect of the WSI representation (p-value < 1.99E-76), classifier type (p-value < 1.73E-21), and AFB threshold (p-value = 0.003) was observed on the BACC. The feature type had no significant effect (p-value = 0.459) on the BACC. WSIs represented by the middle layer, extended focus layer and the z-stack followed by the weighted averaging of AFB probability scores were classified with the average BACC of 58.80%, 68.64%, and 77.28%, respectively. The multilayer WSIs represented by the z-stack with the weighted averaging of AFB probability scores were classified by a Random Forest classifier with the average BACC of 83.32%. Low classification accuracy of WSIs represented by the middle layer suggests that they contain fewer features permitting identification of AFB than the multilayer WSIs. Our results indicate that single-layer acquisition can introduce a bias (sampling error) into the WSI. This bias can be mitigated by the multilayer or the extended focus acquisitions.

iPSC-derived tenocytes seeded on microgrooved 3D printed scaffolds for Achilles tendon regeneration.

Tendons and ligaments have a poor innate healing capacity, yet account for 50% of musculoskeletal injuries in the United States. Full structure and function restoration postinjury remains an unmet clinical need. This study aimed to assess the application of novel three dimensional (3D) printed scaffolds and induced pluripotent stem cell-derived mesenchymal stem cells (iMSCs) overexpressing the transcription factor Scleraxis (SCX, iMSCSCX+ ) as a new strategy for tendon defect repair. The polycaprolactone (PCL) scaffolds were fabricated by extrusion through a patterned nozzle or conventional round nozzle. Scaffolds were seeded with iMSCSCX+ and outcomes were assessed in vitro via gene expression analysis and immunofluorescence. In vivo, rat Achilles tendon defects were repaired with iMSCSCX+ -seeded microgrooved scaffolds, microgrooved scaffolds only, or suture only and assessed via gait, gene expression, biomechanical testing, histology, and immunofluorescence. iMSCSCX+ -seeded on microgrooved scaffolds showed upregulation of tendon markers and increased organization and linearity of cells compared to non-patterned scaffolds in vitro. In vivo gait analysis showed improvement in the Scaffold + iMSCSCX+ -treated group compared to the controls. Tensile testing of the tendons demonstrated improved biomechanical properties of the Scaffold + iMSCSCX+ group compared with the controls. Histology and immunofluorescence demonstrated more regular tissue formation in the Scaffold + iMSCSCX+ group. This study demonstrates the potential of 3D-printed scaffolds with cell-instructive surface topography seeded with iMSCSCX+ as an approach to tendon defect repair. Further studies of cell-scaffold constructs can potentially revolutionize tendon reconstruction by advancing the application of 3D printing-based technologies toward patient-specific therapies that improve healing and functional outcomes at both the cellular and tissue level.

Directing iPSC differentiation into iTenocytes using combined scleraxis overexpression and cyclic loading.

Regenerative therapies for tendon are falling behind other tissues due to the lack of an appropriate and potent cell therapeutic candidate. This study aimed to induce tenogenesis using stable Scleraxis (Scx) overexpression in combination with uniaxial mechanical stretch of iPSC-derived mesenchymal stromal-like cells (iMSCs). Scx is the single direct molecular regulator of tendon differentiation known to date. Bone marrow-derived (BM-)MSCs were used as reference. Scx overexpression alone resulted in significantly higher upregulation of tenogenic markers in iMSCs compared to BM-MSCs. Mechanoregulation is known to be a central element guiding tendon development and healing. Mechanical stimulation combined with Scx overexpression resulted in morphometric and cytoskeleton-related changes, upregulation of early and late tendon markers, and increased extracellular matrix deposition and alignment, and tenomodulin perinuclear localization in iMSCs. Our findings suggest that these cells can be differentiated into tenocytes and might be a better candidate for tendon cell therapy applications than BM-MSCs.

Dynamic Changes in the Extracellular Matrix in Primary, Metastatic, and Recurrent Ovarian Cancers.

Cancer-associated fibroblasts (CAFs) and their extracellular matrix are active participants in cancer progression. While it is known that functionally different subpopulations of CAFs co-exist in ovarian cancer, it is unclear whether certain CAF subsets are enriched during metastatic progression and/or chemotherapy. Using computational image analyses of patient-matched primary high-grade serous ovarian carcinomas, synchronous pre-chemotherapy metastases, and metachronous post-chemotherapy metastases from 42 patients, we documented the dynamic spatiotemporal changes in the extracellular matrix, fibroblasts, epithelial cells, immune cells, and CAF subsets expressing different extracellular matrix components. Among the different CAF subsets, COL11A1+ CAFs were associated with linearized collagen fibers and exhibited the greatest enrichment in pre- and post-chemotherapy metastases compared to matched primary tumors. Although pre- and post-chemotherapy metastases were associated with increased CD8+ T cell infiltration, the infiltrate was not always evenly distributed between the stroma and cancer cells, leading to an increased frequency of the immune-excluded phenotype where the majority of CD8+ T cells are present in the tumor stroma but absent from the tumor parenchyma. Overall, most of the differences in the tumor microenvironment were observed between primary tumors and metastases, while fewer differences were observed between pre- and post-treatment metastases. These data suggest that the tumor microenvironment is largely determined by the primary vs. metastatic location of the tumor while chemotherapy does not have a significant impact on the host microenvironment.

Computational pathology in ovarian cancer.

Histopathologic evaluations of tissue sections are key to diagnosing and managing ovarian cancer. Pathologists empirically assess and integrate visual information, such as cellular density, nuclear atypia, mitotic figures, architectural growth patterns, and higher-order patterns, to determine the tumor type and grade, which guides oncologists in selecting appropriate treatment options. Latent data embedded in pathology slides can be extracted using computational imaging. Computers can analyze digital slide images to simultaneously quantify thousands of features, some of which are visible with a manual microscope, such as nuclear size and shape, while others, such as entropy, eccentricity, and fractal dimensions, are quantitatively beyond the grasp of the human mind. Applications of artificial intelligence and machine learning tools to interpret digital image data provide new opportunities to explore and quantify the spatial organization of tissues, cells, and subcellular structures. In comparison to genomic, epigenomic, transcriptomic, and proteomic patterns, morphologic and spatial patterns are expected to be more informative as quantitative biomarkers of complex and dynamic tumor biology. As computational pathology is not limited to visual data, nuanced subvisual alterations that occur in the seemingly "normal" pre-cancer microenvironment could facilitate research in early cancer detection and prevention. Currently, efforts to maximize the utility of computational pathology are focused on integrating image data with other -omics platforms that lack spatial information, thereby providing a new way to relate the molecular, spatial, and microenvironmental characteristics of cancer. Despite a dire need for improvements in ovarian cancer prevention, early detection, and treatment, the ovarian cancer field has lagged behind other cancers in the application of computational pathology. The intent of this review is to encourage ovarian cancer research teams to apply existing and/or develop additional tools in computational pathology for ovarian cancer and actively contribute to advancing this important field.

Unraveling the Mechanism of the 150-Fold Photocurrent Enhancement in Plasma-Treated 2D TMDs.

Two-dimensional (2D) transition metal dichalcogenides (TMDs) are increasingly investigated for applications such as optoelectronic memories, artificial neurons, sensors, and others that require storing photogenerated signals for an extended period. In this work, we report an environment- and gate voltage-dependent photocurrent modulation method of TMD monolayer-based devices (WS2 and MoS2). To achieve this, we introduce structural defects using mild argon-oxygen plasma treatment. The treatment leads to an extraordinary over 150-fold enhancement of the photocurrent in vacuum along with an increase in the relaxation time. A significant environmental and electrostatic dependence of the photocurrent signal is observed. We claim that the effect is a combined result of atomic vacancy introduction and oxide formation, strengthened by optimal wavelength choice for the modified surface. We believe that this work contributes to paving the way for tunable 2D TMD optoelectronic applications.

Efficient automatic 3D segmentation of cell nuclei for high-content screening.

BACKGROUND: High-content screening (HCS) is a pre-clinical approach for the assessment of drug efficacy. On modern platforms, it involves fluorescent image capture using three-dimensional (3D) scanning microscopy. Segmentation of cell nuclei in 3D images is an essential prerequisite to quantify captured fluorescence in cells for screening. However, this segmentation is challenging due to variabilities in cell confluency, drug-induced alterations in cell morphology, and gradual degradation of fluorescence with the depth of scanning. Despite advances in algorithms for segmenting nuclei for HCS, robust 3D methods that are insensitive to these conditions are still lacking. RESULTS: We have developed an algorithm which first generates a 3D nuclear mask in the original images. Next, an iterative 3D marker-controlled watershed segmentation is applied to downsized images to segment adjacent nuclei under the mask. In the last step, borders of segmented nuclei are adjusted in the original images based on local nucleus and background intensities. The method was developed using a set of 10 3D images. Extensive tests on a separate set of 27 3D images containing 2,367 nuclei demonstrated that our method, in comparison with 6 reference methods, achieved the highest precision (PR = 0.97), recall (RE = 0.88) and F1-score (F1 = 0.93) of nuclei detection. The Jaccard index (JI = 0.83), which reflects the accuracy of nuclei delineation, was similar to that yielded by all reference approaches. Our method was on average more than twice as fast as the reference method that produced the best results. Additional tests carried out on three stacked 3D images comprising heterogenous nuclei yielded average PR = 0.96, RE = 0.84, F1 = 0.89, and JI = 0.80. CONCLUSIONS: The high-performance metrics yielded by the proposed approach suggest that it can be used to reliably delineate nuclei in 3D images of monolayered and stacked cells exposed to cytotoxic drugs.

Focal Serous Tubal Intra-Epithelial Carcinoma Lesions Are Associated With Global Changes in the Fallopian Tube Epithelia and Stroma.

Objective: Serous tubal intra-epithelial carcinoma (STIC) lesions are thought to be precursors to high-grade serous ovarian cancer (HGSOC), but HGSOC is not always accompanied by STIC. Our study was designed to determine if there are global visual and subvisual microenvironmental differences between fallopian tubes with and without STIC lesions. Methods: Computational image analyses were used to identify potential morphometric and topologic differences in stromal and epithelial cells in samples from three age-matched groups of fallopian tubes. The Benign group comprised normal fallopian tubes from women with benign conditions while the STIC and NoSTIC groups consisted of fallopian tubes from women with HGSOC, with and without STIC lesions, respectively. For the morphometric feature extraction and analysis of the stromal architecture, the image tiles in the STIC group were further divided into the stroma away from the STIC (AwaySTIC) and the stroma near the STIC (NearSTIC). QuPath software was used to identify and quantitate secretory and ciliated epithelial cells. A secretory cell expansion (SCE) or a ciliated cell expansion (CCE) was defined as a monolayered contiguous run of >10 secretory or ciliated cells uninterrupted by the other cell type. Results: Image analyses of the tubal stroma revealed gradual architectural differences from the Benign to NoSTIC to AwaySTIC to NearSTIC groups. In the epithelial topology analysis, the relative number of SCE and the average number of cells within SCE were higher in the STIC group than in the Benign and NoSTIC groups. In addition, aging was associated with an increased relative number of SCE and a decreased relative number of CCE. ROC analysis determined that an average of 15 cells within SCE was the optimal cutoff value indicating the presence of a STIC lesion in the tubal epithelium. Conclusions: Our findings suggest that global stromal alterations and age-associated reorganization of tubal secretory and ciliated cells are associated with STIC lesions. Further studies will need to determine if these alterations precede STIC lesions and provide permissible conditions for the formation of STIC.

Receptor-interacting protein kinase 2 (RIPK2) stabilizes c-Myc and is a therapeutic target in prostate cancer metastasis.

Despite progress in prostate cancer (PC) therapeutics, distant metastasis remains a major cause of morbidity and mortality from PC. Thus, there is growing recognition that preventing or delaying PC metastasis holds great potential for substantially improving patient outcomes. Here we show receptor-interacting protein kinase 2 (RIPK2) is a clinically actionable target for inhibiting PC metastasis. RIPK2 is amplified/gained in ~65% of lethal metastatic castration-resistant PC. Its overexpression is associated with disease progression and poor prognosis, and its genetic knockout substantially reduces PC metastasis. Multi-level proteomics analyses reveal that RIPK2 strongly regulates the stability and activity of c-Myc (a driver of metastasis), largely via binding to and activating mitogen-activated protein kinase kinase 7 (MKK7), which we identify as a direct c-Myc-S62 kinase. RIPK2 inhibition by preclinical and clinical drugs inactivates the noncanonical RIPK2/MKK7/c-Myc pathway and effectively impairs PC metastatic outgrowth. These results support targeting RIPK2 signaling to extend metastasis-free and overall survival.

Modelling and Characterisation of Residual Stress of SiC-Ti3C2Tx MXene Composites Sintered via Spark Plasma Sintering Method.

This article presents an attempt to determine the effect of the MXene phase addition and its decomposition during sintering with the use of the spark plasma sintering method on mechanical properties and residual stress of silicon carbide based composites. For this purpose, the unreinforced silicon carbide sinter and the silicon carbide composite with the addition of 2 wt.% of Ti3C2Tx were tested. The results showed a significant increase of fracture toughness and hardness for composite, respectively 36% and 13%. The numerical study involving this novel method of modelling shows the presence of a complex state of stress in the material, which is related to the anisotropic properties of graphitic carbon structures formed during sintering. An attempt to determine the actual values of residual stress in the tested materials using Raman spectroscopy was also made. These tests showed a good correlation with the constructed numerical model and confirmed the presence of a complex state of residual stress.

A U-Net based framework to quantify glomerulosclerosis in digitized PAS and H&E stained human tissues.

Reliable counting of glomeruli and evaluation of glomerulosclerosis in renal specimens are essential steps to assess morphological changes in kidney and identify individuals requiring treatment. Because microscopic identification of sclerosed glomeruli performed under the microscope is labor intensive, we developed a deep learning (DL) approach to identify and classify glomeruli as normal or sclerosed in digital whole slide images (WSIs). The segmentation and classification of glomeruli was performed by the U-Net model. Subsequently, glomerular classifications were refined based on glomerular histomorphometry. The U-Net model was trained using patches from Periodic Acid-Schiff (PAS) stained WSIs (n=31) from the AIDPATH - a multi-center dataset, and then tested on an independent set of WSIs (n=20) including PAS (n=6), and hematoxylin and eosin (H&E) stained WSIs (n=14) from four other institutions. The training and test WSIs were obtained from formalin fixed and paraffin embedded blocks with of human kidney specimens each presenting various proportions of normal and sclerosed glomeruli. In the PAS stained WSIs, normal and sclerosed glomeruli were respectively classified with the F1-score of 97.5% and 68.8%. In the H&E stained WSIs, the F1-scores of 90.8% and 78.1% were achieved. Regardless the tissue staining, the glomeruli in the test WSIs were classified with the F1-score of 94.5% (n=923, normal) and 76.8% for (n=261, sclerosed). These results demonstrate for the first time that a framework based on the U-Net model trained with glomerular patches from PAS stained WSIs can reliably segment and classify normal and sclerosed glomeruli in PAS and also H&E stained WSIs. Our approach yielded higher accuracy of glomerular classifications than some of the recently published methods. Additionally, our test set of images with ground truth is publicly available.

Study of Structural and Optoelectronic Properties of Thin Films Made of a Few Layered WS2 Flakes.

We report a surfactant-free exfoliation method of WS2 flakes combined with a vacuum filtration method to fabricate thin (<50 nm) WS2 films, that can be transferred on any arbitrary substrate. Films are composed of thin (<4 nm) single flakes, forming a large size uniform film, verified by AFM and SEM. Using statistical phonons investigation, we demonstrate structural quality and uniformity of the film sample and we provide first-order temperature coefficient χ, which shows linear dependence over 300-450 K temperature range. Electrical measurements show film sheet resistance RS = 48 MΩ/Υ and also reveal two energy band gaps related to the intrinsic architecture of the thin film. Finally, we show that optical transmission/absorption is rich above the bandgap exhibiting several excitonic resonances, and nearly feature-less below the bandgap.

Substrate-Induced Variances in Morphological and Structural Properties of MoS2 Grown by Chemical Vapor Deposition on Epitaxial Graphene and SiO2.

In this work, we report the impact of substrate type on the morphological and structural properties of molybdenum disulfide (MoS2) grown by chemical vapor deposition (CVD). MoS2 synthesized on a three-dimensional (3D) substrate, that is, SiO2, in response to the change of the thermodynamic conditions yielded different grain morphologies, including triangles, truncated triangles, and circles. Simultaneously, MoS2 on graphene is highly immune to the modifications of the growth conditions, forming triangular crystals only. We explain the differences between MoS2 on SiO2 and graphene by the different surface diffusion mechanisms, namely, hopping and gas-molecule-collision-like mechanisms, respectively. As a result, we observe the formation of thermodynamically favorable nuclei shapes on graphene, while on SiO2, a full spectrum of domain shapes can be achieved. Additionally, graphene withstands the growth process well, with only slight changes in strain and doping. Furthermore, by the application of graphene as a growth substrate, we realize van der Waals epitaxy and achieve strain-free growth, as suggested by the photoluminescence (PL) studies. We indicate that PL, contrary to Raman spectroscopy, enables us to arbitrarily determine the strain levels in MoS2.

A CNN-based active learning framework to identify mycobacteria in digitized Ziehl-Neelsen stained human tissues.

Tuberculosis is the most common mycobacterial disease that affects humans worldwide. Rapid and reliable diagnosis of mycobacteria is crucial to identify infected individuals, to initiate and monitor treatment and to minimize or prevent transmission. Microscopic identification of acid-fast mycobacteria (AFB) in tissue sections is usually accomplished by examining Ziehl-Neelsen (ZN) stained slides in which AFB appear bright red against the blue background. Because the ZN-stained slides require time consuming and meticulous screening by an experienced pathologist, our team developed a machine learning pipeline to classify digitized ZN-stained slides as AFB-positive or AFB-negative. The pipeline includes two convolutional neural network (CNN) models to recognize tiles containing AFB, and a logistic regression (LR) model to classify slides based on features from AFB-probability maps assembled from the CNN tile-based classification results. The first CNN was trained using tiles from 6 AFB-positive and 8 AFB-negative slides, and the second CNN was trained using the initial tile set expanded by additional tiles from 19 AFB-negative slides selected within an active learning framework. When evaluated on a separate set of tiles, the two CNNs yielded F1 scores of 99.03% and 98.75%, respectively, and were used to classify tiles in a separate set of 134 slides (46 AFB-positive and 88 AFB-negative). The classification yielded two AFB-probability maps, one for each CNN. The LR model was then 10-fold cross-validated using the average of feature vectors extracted from the AFB-probability maps generated by each CNN. The feature vector consisted of seven features of the AFB-probability map histogram and the positive tile rate (PTR). The sensitivity (87.13%), specificity (87.62%) and F1 (80.18%) achieved by this model were superior to the baseline performance of PTR-based separation of slides that yielded F1 scores of 73.13% and 66.67% in the AFB-probability maps outputted by the CNN trained within the active learning framework and the CNN trained only on the initial set of slides, respectively. Our CNNs outperformed several recently published models for AFB detection. Active learning induced robust learning of features by the CNN and led to improved LR classification performance of slides. In the 52 AFB-positive slides used in the pipeline development, the AFB were infrequent, predominantly single and only rarely found in small clusters. Our pipeline can classify slides and visualize suspected AFB-positive areas in each slide, and thus potentially facilitate evaluation of ZN-stained tissue sections for AFB.

Predicting Metastasis Risk in Pancreatic Neuroendocrine Tumors Using Deep Learning Image Analysis.

BACKGROUND: The prognosis of patients with pancreatic neuroendocrine tumors (PanNET), the second most common type of pancreatic cancer, varies significantly, and up to 15% of patients develop metastasis. Although certain morphological characteristics of PanNETs have been associated with patient outcome, there are no available morphology-based prognostic markers. Given that current clinical histopathology markers are unable to identify high-risk PanNET patients, the development of accurate prognostic biomarkers is needed. Here, we describe a novel machine learning, multiclassification pipeline to predict the risk of metastasis using morphological information from whole tissue slides. METHODS: Digital images from surgically resected tissues from 89 PanNET patients were used. Pathologist-annotated regions were extracted to train a convolutional neural network (CNN) to identify tiles consisting of PanNET, stroma, normal pancreas parenchyma, and fat. Computationally annotated cancer or stroma tiles and patient metastasis status were used to train CNN to calculate a region based metastatic risk score. Aggregation of the metastatic probability scores across the slide was performed to predict the risk of metastasis. RESULTS: The ability of CNN to discriminate different tissues was high (per-tile accuracy >95%; whole slide cancer regions Jaccard index = 79%). Cancer and stromal tiles with high evaluated probability provided F1 scores of 0.82 and 0.69, respectively, when we compared tissues from patients who developed metastasis and those who did not. The final model identified low-risk (n = 76) and high-risk (n = 13) patients, as well as predicted metastasis-free survival (hazard ratio: 4.71) after adjusting for common clinicopathological variables, especially in grade I/II patients. CONCLUSION: Using slides from surgically resected PanNETs, our novel, multiclassification, deep learning pipeline was able to predict the risk of metastasis in PanNET patients. Our results suggest the presence of prognostic morphological patterns in PanNET tissues, and that these patterns may help guide clinical decision making.