Ground Truth Research Articles

Bcubed revisited: elements like me

BCubed is a mathematically clean, elegant and intuitively well behaved external performance metric for clustering tasks. BCubed compares a predicted clustering to a known ground truth clustering through elementwise precision and recall scores. For each element, the predicted and ground truth clusters containing the element are compared, and the mean over all elements is taken. We argue that BCubed overestimates performance, for the intuitive reason that the clustering gets credit for putting an element into its own cluster. This is repaired, and we investigate the repaired version, called “Elements Like Me (ELM)”. We extensively evaluate ELM from both a theoretical and empirical perspective, and conclude that it retains all of its positive properties, and yields a minimum zero score when it should. Synthetic experiments show that ELM can produce different rankings of predicted clusterings when compared to BCubed, and that the ELM scores are distributed with lower mean and a larger variance than BCubed.

Registration of Interferometric DEM by Deep Artificial Neural Networks Using GPS Control Points’ Coordinates as Network Target

The Shuttle Radar Topography Mission (SRTM) satellite’s digital elevation model (DEM) is an important tool for studying topographic features on a medium-spacing scale. Data were collected and processed using the satellite’s orbital and navigation parameters with selected global GPS stations for verification. Distortion may be expressed by surveying measurements, such as position, distance, area, and shape. This study focuses on this distortion and proposes a new registration method to reduce its effect. Because of generality, the purpose shapes were excluded from this study. The proposed registration method depends on precise GPS control points that act as the ground truth for describing the considered surveying measurements. The processing was carried out using deep artificial neural networks (DANN) to produce a new registered DEM. A comparison was made between the original DEM and the new one, focusing on the selected surveying measurements. Another comparison was made between the GPS coordinates and SRTM polynomials to determine the potential of the proposed system. Some statistical investigations were applied to determine the level of significance of the distortion in each surveying measurement. The study shows that the distortion is highly significant; therefore, the proposed registration method is recommended to fix the distortion.

Typicality- and instance-dependent label noise-combating: a novel framework for simulating and combating real-world noisy labels for endoscopic polyp classification

Learning with noisy labels aims to train neural networks with noisy labels. Current models handle instance-independent label noise (IIN) well; however, they fall short with real-world noise. In medical image classification, atypical samples frequently receive incorrect labels, rendering instance-dependent label noise (IDN) an accurate representation of real-world scenarios. However, the current IDN approaches fail to consider the typicality of samples, which hampers their ability to address real-world label noise effectively. To alleviate the issues, we introduce typicality- and instance-dependent label noise (TIDN) to simulate real-world noise and establish a TIDN-combating framework to combat label noise. Specifically, we use the sample’s distance to decision boundaries in the feature space to represent typicality. The TIDN is then generated according to typicality. We establish a TIDN-attention module to combat label noise and learn the transition matrix from latent ground truth to the observed noisy labels. A recursive algorithm that enables the network to make correct predictions with corrections from the learned transition matrix is proposed. Our experiments demonstrate that the TIDN simulates real-world noise more closely than the existing IIN and IDN. Furthermore, the TIDN-combating framework demonstrates superior classification performance when training with simulated TIDN and actual real-world noise.

Automated Porosity Characterization for Aluminum Die Casting Materials Using X-ray Radiography, Synthetic X-ray Data Augmentation by Simulation, and Machine Learning.

Detection and characterization of hidden defects, impurities, and damages in homogeneous materials like aluminum die casting materials, as well as composite materials like Fiber-Metal Laminates (FML), is still a challenge. This work discusses methods and challenges in data-driven modeling of automated damage and defect detectors using measured X-ray single- and multi-projection images. Three main issues are identified: Data and feature variance, data feature labeling (for supervised machine learning), and the missing ground truth. It will be shown that simulation of synthetic measuring data can deliver a ground truth dataset and accurate labeling for data-driven modeling, but it cannot be used directly to predict defects in manufacturing processes. Noise has a significant impact on the feature detection and will be discussed. Data-driven feature detectors are implemented with semantic pixel Convolutional Neural Networks. Experimental data are measured with different devices: A low-quality and low-cost (Low-Q) X-ray radiography, a typical industrial mid-quality X-ray radiography and Computed Tomography (CT) system, and a state-of-the-art high-quality μ-CT device. The goals of this work are the training of robust and generalized data-driven ML feature detectors with synthetic data only and the transition from CT to single-projection radiography imaging and analysis. Although, as the title implies, the primary task is pore characterization in aluminum high-pressure die-cast materials, but the methods and results are not limited to this use case.

230 Using phenotypes with uncertain identification leveraging computer vision models

Abstract There has been an increase in high throughput phenotyping using sensors in livestock farming. The major obstacle when analyzing phenotypic data in studies using cameras is identifying the animals using computer vision. To overcome this problem, an analysis method that uses keypoint detection in conjunction with classification algorithms can identify animals in images. However, these algorithms generally output a prediction probability for different classes, reflecting uncertainty. In most practices, the highest probability will be assigned as the putative identification (ID). In some cases, this may introduce incorrect identification and affect downstream decisions. Here, we propose an analytical approach to incorporate all the uncertainty in the putative ID distribution into downstream inferences. We used a publicly available keypoint detection dataset of 17 keypoints detected in 3,226 images from 30 diverse thoroughbred horses. We performed feature engineering with Linear Discriminant Analysis (LDA) using 5-fold cross-validation to build horse classifiers. The best model selected had an accuracy of 88.12% to identify animals in the testing set. Afterward, we performed Monte Carlo simulations to add a phenotype that was assumed independent of the image keypoints. This generated a set of plasmodes to evaluate methods that accounted for ID uncertainty. Each simulation included different means for two groups of 15 horses. In total nine simulation scenarios were considered including varying effect sizes of group mean difference and various repeatability values (horse variance to total variance ratio). We analyzed the simulated data with three different mixed models: 1) Fixed and random effects were based on using ground truth assignment of ID, 2) Fixed and random effects were based on using the putative ID with the highest posterior probability from the best classifier, 3) Finally, prediction probabilities were used to account for putative ID uncertainty. As expected, the model that used ground truth ID had the highest power to detect group differences, and it had the lowest root mean square error (rMSE) to estimate variance components across all scenarios. The linear model that used highest posterior probability putative ID, showed the lowest power and highest rMSE. For instance, as shown in Table 1, the mean rMSE value for group mean differences in the model using ground truth is 0.01 and the model using highest posterior probabilities is 0.13. The model that incorporated ID uncertainty had intermediate properties, showing a rMSE of 0.03 for group differences. Overall, incorporating uncertainty in Animal ID results in improved performance in ultimate test statistics used in animal studies. This is very important in breeding and precision livestock farming contexts where animal identification is challenging. We will further investigate this.

Tracking and co-location of global point clouds for large-area indoor environments

Extended reality (XR) experiences are on the verge of becoming widely adopted in diverse application domains. An essential part of the technology is accurate tracking and localization of the headset to create an immersive experience. A subset of the applications require perfect co-location between the real and the virtual world, where virtual objects are aligned with real-world counterparts. Current headsets support co-location for small areas, but suffer from drift when scaling up to larger ones such as buildings or factories. This paper proposes tools and solutions for this challenge by splitting up the simultaneous localization and mapping (SLAM) into separate mapping and localization stages. In the pre-processing stage, a feature map is built for the entire tracking area. A global optimizer is applied to correct the deformations caused by drift, guided by a sparse set of ground truth markers in the point cloud of a laser scan. Optionally, further refinement is applied by matching features between the ground truth keyframe images and their rendered-out SLAM estimates of the point cloud. In the second, real-time stage, the rectified feature map is used to perform localization and sensor fusion between the global tracking and the headset. The results show that the approach achieves robust co-location between the virtual and the real 3D environment for large and complex tracking environments.

Improved quantification in CEST-MRI by joint spatial total generalized variation.

In this work, the use of joint Total Generalized Variation (TGV) regularization to improve Multipool-Lorentzian fitting of chemical exchange saturation transfer (CEST) Spectra in terms of stability and parameter signal-to-noise ratio (SNR) was investigated. The joint TGV term was integrated into the nonlinear parameter fitting problem. To increase convergence and weight the gradients, preconditioning using a voxel-wise singular value decomposition was applied to the problem, which was then solved using the iteratively regularized Gauss-Newton method combined with a Primal-Dual splitting algorithm. The TGV method was evaluated on simulated numerical phantoms, 3T phantom data and 7T in vivo data with respect to systematic errors and robustness. Three reference methods were also implemented: The standard nonlinear fitting, a method using a nonlocal-means filter for denoising and the pyramid scheme, which uses downsampled images to acquire accurate start values. The proposed regularized fitting method showed significantly improved robustness (compared to the reference methods). In testing, over a range of SNR values the TGV fit outperformed the other methods and showed accurate results even for large amounts of added noise. Parameter values found were closer or comparable to the ground truth. For in vivo datasets, the added regularization increased the parameter map SNR and prevented instabilities. The proposed fitting method using TGV regularization leads to improved results over a range of different data-sets and noise levels. Furthermore, it can be applied to all Z-spectrum data, with different amounts of pools, where the improved SNR and stability can increase diagnostic confidence.

Characterizing the voxel-based approaches in radioembolization dosimetry with reDoseMC.

Yttrium-90 ( ) represents the primary radioisotope used in radioembolization procedures, while holmium-166 ( ) is hypothesized to serve as a viable substitute for due to its comparable therapeutic potential and improved quantitative imaging. Voxel-based dosimetry for these radioisotopes relies on activity images obtained through PET or SPECT and dosimetry methods, including the voxel S-value (VSV) and the local deposition method (LDM). However, the evaluation of the accuracy of absorbed dose calculations has been limited by the use of non-ideal reference standards and investigations restricted to the liver. The objective of this study was to expand upon these dosimetry characterizations by investigating the impact of image resolutions, voxel sizes, target volumes, and tissue materials on the accuracy of and dosimetry techniques. A specialized radiopharmaceutical dosimetry software called reDoseMC was developed using the Geant4 Monte Carlo toolkit and validated by benchmarking the generated kernels with published data. The decay spectra of both and were also compared. Multiple VSV kernels were generated for the liver, lungs, soft tissue, and bone for isotropic voxel sizes of 1 mm, 2 mm, and 4 mm. Three theoretical phantom setups were created with 20 or 40mm activity and mass density inserts for the same three voxel sizes. To replicate the limited spatial resolutions present in PET and SPECT images, image resolutions were modeled using a 3D Gaussian kernel with a Full Width at Half Maximum (FWHM) ranging from 0 to 16mm and with no added noise. The VSV and LDM dosimetry methods were evaluated by characterizing their respective kernels and analyzing their absorbed dose estimates calculated on theoretical phantoms. The ground truth for these estimations was calculated usingreDoseMC. The decay spectra obtained through reDoseMC showed less than a 1% difference when compared to previously published experimental data for energies below 1.9MeV in the case of and less than 1% for energies below 1.5MeV for . Additionally, the validation kernels for VSV exhibited results similar to those found in published Monte Carlo codes, with source dose depositions having less than a 3% error margin. Resolution thresholds ( ), defined as resolutions that resulted in similar dose estimates between the LDM and VSV methods, were observed for . They were 1.5mm for bone, 2.5 mm for soft tissue and liver, and 8.5 mm for lungs. For , the accuracy of absorbed dose deposition was found to be dependent on the contributions of absorbed dose from photons. Volume errors due to variations in voxel size impacted the final dose estimates. Larger target volumes yielded more accurate mean doses than smaller volumes. For both radioisotopes, the radial dose profiles for the VSV and LDM approximated but never matched the referencestandard. reDoseMC was developed and validated for radiopharmaceutical dosimetry. The accuracy of voxel-based dosimetry was found to vary widely with changes in image resolutions, voxel sizes, chosen target volumes, and tissue material; hence, the standardization of dosimetry protocols was found to be of great importance for comparable dosimetryanalysis.

Artificial neural network for enhancing signal-to-noise ratio and contrast in photothermal optical coherence tomography

Optical coherence tomography (OCT) is a medical imaging method that generates micron-resolution 3D volumetric images of tissues in-vivo. Photothermal (PT)-OCT is a functional extension of OCT with the potential to provide depth-resolved molecular information complementary to the OCT structural images. PT-OCT typically requires long acquisition times to measure small fluctuations in the OCT phase signal. Here, we use machine learning with a neural network to infer the amplitude of the photothermal phase modulation from a short signal trace, trained in a supervised fashion with the ground truth signal obtained by conventional reconstruction of the PT-OCT signal from a longer acquisition trace. Results from phantom and tissue studies show that the developed network improves signal to noise ratio (SNR) and contrast, enabling PT-OCT imaging with short acquisition times and without any hardware modification to the PT-OCT system. The developed network removes one of the key barriers in translation of PT-OCT (i.e., long acquisition time) to the clinic.

AI-based diagnosis of chronic obstructive pulmonary disease from low-dose CT images

Background: Chronic obstructive pulmonary disease (COPD) is a group of diseases characterized by airflow blockage. It is one of the leading causes of global mortality and is primarily attributed to smoking. COPD patients are usually diagnosed by spirometry test. Although regarded as the gold standard for COPD diagnosis, the spirometry test carries contraindications, thus prompting the development of low-dose computed tomography (low-dose CT) scan as an alternative for COPD screening. However, a practical limitation of diagnosing COPD from CT images is its reliance on the expertise of a skilled radiologist. Objective: To address this limitation, we aimed to develop a deep-learning model for the automated classification of COPD and non-COPD from low-dose CT images. Materials and methods: We examined the potential of a convolutional neural network for identifying COPD. Our dataset consisted of 10,000 low-dose CT images obtained from a lung cancer screening program, involving both ex-smokers and current smokers deemed at high risk of lung cancer. Spirometry data served as the ground truth for defining COPD. We used 90% of the datasets for training and 10% for testing. Results: Our developed model achieved notable performance metrics: an area under the receiver operating characteristic curve (AUC) of 0.97, an accuracy of 0.89, a precision of 0.85, a recall of 0.96, and an F1-score of 0.90. Conclusion: Our study demonstrates the potential of deep learning models to augment clinical assessments and improve the diagnosis of COPD, thereby enhancing diagnostic accuracy and efficiency. The findings suggest the feasibility of integrating this technology into routine lung cancer screening programs for COPD detection.

Robformer: A robust decomposition transformer for long-term time series forecasting

Transformer-based forecasting methods have been widely applied to forecast long-term multivariate time series, which achieves significant improvements on extending the forecasting time. However, their performance can degenerate terribly when abrupt trend shift and seasonal fluctuation arise in long-term time series. Hence, we identify two bottlenecks of previous Transformers architecture: (1) the robustless decomposition module and (2) trend shifting problem. These result in a different distribution between the trend prediction and ground truth in the long-term multivariate series forecasting. Towards these bottlenecks, we design Robformer as a novel decomposition-based Transformer, which consists of three new inner module to enhance the predictability of Transformers. Concretely, we renew the decomposition module and add a seasonal component adjustment module to tackle the unstationarized series. Further, we propose a novel inner trend forecasting architecture inspired by polynomial fitting method, which outperforms previous design in accuracy and robustness. Our empirical studies show that Robformer can achieve 17% and 10% relative improvements than state-of-the-art Autoformer and FEDformer baselines under the fair long-term multivariate setting on six benchmarks, covering five mainstream time series forecasting applications: energy, economics, traffic, weather, and disease. The code will be released upon publication.

An Artificial Intelligence model for implant segmentation on periapical radiographs.

To segment dental implants on PA radiographs using a Deep Learning (DL) algorithm. To compare the performance of the algorithm relative to ground truth determined by the human annotator. Three hundred PA radiographs were retrieved from the radiographic database and consequently annotated to label implants as well as teeth on the LabelMe annotation software. The dataset was augmented to increase the number of images in the training data and a total of 1294 images were used to train, validate and test the DL algorithm. An untrained U-net was downloaded and trained on the annotated dataset to allow detection of implants using polygons on PA radiographs. A total of one hundred and thirty unseen images were run through the trained U-net to determine its ability to segment implants on PA radiographs. The performance metrics are as follows: accuracy of 93.8%, precision of 90%, recall of 83%, F-1 score of 86%, Intersection over Union of 86.4% and loss = 21%. The trained DL algorithm segmented implants on PA radiographs with high performance similar to that of the humans who labelled the images forming the ground truth.

Seagrass mapping across the coast of Tanzania

Seagrasses are marine flowering plants found in both temperate and tropical coastal environments, that offer vital ecosystem services and adaptations. However, due to their proximity to the shore, they are heavily impacted by human activities and climate change, endangering seagrasses and their services. To manage seagrass meadows effectively and sustainably, detailed monitoring and data on distribution and extent are essential. Unfortunately, past under-prioritization of reliable baseline data impedes understanding of ecosystem dynamics and rapid response to declines. This study employs satellite imagery, coupled with ground truthing, to comprehensively report seagrass coverage along Tanzania's mainland coast and Zanzibar islands. By utilizing machine-learning technologies, the study achieves 84.2% accuracy, and found that Tanzania has 2817 km2 of seagrass coverage, which is five times higher than previously reported. In addition, the results also provide coverage of additional benthic habitats, such as corals, rocks/rubble, and sand. This study provides the first comprehensive overview of seagrass coverage in Tanzania and will function as a valuable tool and baseline for understanding and managing these habitats. The findings suggest that seagrasses in Tanzania are still relatively abundant, but due to a number of threats, including pollution, coastal development, and climate change, these habitats require improvement in management strategies including regular monitoring to understand their status and trends in their area coverage. The study highlights the significance of remote sensing and machine learning in mapping and monitoring coastal habitats across extensive geographical areas. We aspire for the presented results to be utilized among stakeholders and researchers to develop more effective conservation measures for seagrasses in Tanzania and the Western Indian Ocean.

Accelerate microstructure evolution simulation using graph neural networks with adaptive spatiotemporal resolution

Surrogate models driven by sizeable datasets and scientific machine-learning methods have emerged as an attractive microstructure simulation tool with the potential to deliver predictive microstructure evolution dynamics with huge savings in computational costs. Taking 2D and 3D grain growth simulations as an example, we present a completely overhauled computational framework based on graph neural networks with not only excellent agreement to both the ground truth phase-field methods and theoretical predictions, but enhanced accuracy and efficiency compared to previous works based on convolutional neural networks. These improvements can be attributed to the graph representation, both improved predictive power and a more flexible data structure amenable to adaptive mesh refinement. As the simulated microstructures coarsen, our method can adaptively adopt remeshed grids and larger timesteps to achieve further speedup. The data-to-model pipeline with training procedures together with the source codes are provided.

Detecting the symptoms of Parkinson’s disease with non-standard video

BackgroundNeurodegenerative diseases, such as Parkinson’s disease (PD), necessitate frequent clinical visits and monitoring to identify changes in motor symptoms and provide appropriate care. By applying machine learning techniques to video data, automated video analysis has emerged as a promising approach to track and analyze motor symptoms, which could facilitate more timely intervention. However, existing solutions often rely on specialized equipment and recording procedures, which limits their usability in unstructured settings like the home. In this study, we developed a method to detect PD symptoms from unstructured videos of clinical assessments, without the need for specialized equipment or recording procedures.MethodsTwenty-eight individuals with Parkinson’s disease completed a video-recorded motor examination that included the finger-to-nose and hand pronation-supination tasks. Clinical staff provided ground truth scores for the level of Parkinsonian symptoms present. For each video, we used a pre-existing model called PIXIE to measure the location of several joints on the person’s body and quantify how they were moving. Features derived from the joint angles and trajectories, designed to be robust to recording angle, were then used to train two types of machine-learning classifiers (random forests and support vector machines) to detect the presence of PD symptoms.ResultsThe support vector machine trained on the finger-to-nose task had an F1 score of 0.93 while the random forest trained on the same task yielded an F1 score of 0.85. The support vector machine and random forest trained on the hand pronation-supination task had F1 scores of 0.20 and 0.33, respectively.ConclusionThese results demonstrate the feasibility of developing video analysis tools to track motor symptoms across variable perspectives. These tools do not work equally well for all tasks, however. This technology has the potential to overcome barriers to access for many individuals with degenerative neurological diseases like PD, providing them with a more convenient and timely method to monitor symptom progression, without requiring a structured video recording procedure. Ultimately, more frequent and objective home assessments of motor function could enable more precise telehealth optimization of interventions to improve clinical outcomes inside and outside of the clinic.

Graph Sampling for Map Comparison

Comparing two road maps is a basic operation that arises in a variety of situations. A map comparison method that is commonly used, mainly in the context of comparing reconstructed maps to ground truth maps, is based on graph sampling . The essential idea is to first compute a set of point samples on each map, and then to match pairs of samples—one from each map—in a one-to-one fashion. For deciding whether two samples can be matched, different criteria, e.g., based on distance or orientation, can be used. The total number of matched pairs gives a measure of how similar the maps are. Since the work of Biagioni and Eriksson [11, 12], graph sampling methods have become widely used. However, there are different ways to implement each of the steps, which can lead to significant differences in the results. This means that conclusions drawn from different studies that seemingly use the same comparison method, cannot necessarily be compared. In this work we present a unified approach to graph sampling for map comparison. We present the method in full generality, discussing the main decisions involved in its implementation. In particular, we point out the importance of the sampling method (GEO vs. TOPO) and that of the matching definition, discussing the main options used, and proposing alternatives for both key steps. We experimentally evaluate the different sampling and matching options considered on map datasets and reconstructed maps. Furthermore, we provide a code base and an interactive visualization tool to set a standard for future evaluations in the field of map construction and map comparison.

SAR deep learning sea ice retrieval trained with airborne laser scanner measurements from the MOSAiC expedition

Abstract. Automated sea ice charting from synthetic aperture radar (SAR) has been researched for more than a decade, and we are still not close to unlocking the full potential of automated solutions in terms of resolution and accuracy. The central complications arise from ground truth data not being readily available in the polar regions. In this paper, we build a data set from 20 near-coincident x-band SAR acquisitions and as many airborne laser scanner (ALS) measurements from the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC), between October and May. This data set is then used to assess the accuracy and robustness of five machine-learning-based approaches by deriving classes from the freeboard, surface roughness (standard deviation at 0.5 m correlation length) and reflectance. It is shown that there is only a weak correlation between the radar backscatter and the sea ice topography. Accuracies between 44 % and 66 % and robustness between 71 % and 83 % give a realistic insight into the performance of modern convolutional neural network architectures across a range of ice conditions over 8 months. It also marks the first time algorithms have been trained entirely with labels from coincident measurements, allowing for a probabilistic class retrieval. The results show that segmentation models able to learn from the class distribution perform significantly better than pixel-wise classification approaches by nearly 20 % accuracy on average.

Abstract PS03-09: Age-related normal breast tissue features differ in women with germline BRCA1/2 mutations

Abstract Ageing, a breast cancer risk factor, can be reflected in histologically normal breast tissue (NBT). However, the relationship between age-related histological features and cancer risk is not fully understood. We propose that biological and microscopic features can be identified in seemingly histologically NBT in gBRCA1/2m carriers that are indicative of the earliest tissue changes of breast cancer, potentially as a result of accelerated tissue ageing. To study NBT, we have put in place a unique and scalable repository, named OASIS, storing currently >2,000 digitised whole slide images (WSI) of H&E-stained NBT from individuals with long clinical follow-up. These tissues are from individuals with a spectrum of risk for developing breast cancer and derived from reduction surgery from non-gBRCA1/2m carriers, risk reducing masectomies, and from contralateral and peri-tumoural NBT from gBRCA1/2m carriers and women with breast cancer. Across 6 NBT resources, 70 selected WSIs were manually annotated for epithelial cells, fibrous stroma and adipocytes, resembling the "ground truth". Extensive comparisons of different tile sizes, tile overlapping ratios, stain normalisation techniques and deep learning (DL) feature extractors were conducted to implement a Tissue-Classifier, which achieved 95% accuracy in 3-fold cross-validation. Based on the MobileNet architecture, a DL-framework achieved a sensitivity of 81.2% and specificity of 81.8% (84.7% AUC) in predicting Tissue-Age based on discernible histological patterns in healthy women. In gBRCA1/2m carriers, we discovered features of accelerated tissue ageing (e.g. degree of lobule involution), indicating that the biological NBT age differs from their chronological age. Our DL-based framework robustly captured histological patterns to predict tissue components and age. The global organisation of these patterns showed variations in WSIs of women with different risks of developing breast cancers, which may provide new insights into premalignant alterations in NBT. Citation Format: Anita Grigoriadis, Mario Parreno-Centeno, Siyuan Chen, Gregory Verghese, Esther Lips, Cheryl Gillett, Louise Jones, Sarah Pinder. Age-related normal breast tissue features differ in women with germline BRCA1/2 mutations [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PS03-09.

Deep learning-based prediction of thermal residual stress and melt pool characteristics in laser-irradiated carbon steel

Residual stress is a crucial factor that must be managed in laser manufacturing processes. Neglecting it can result in product failure or unsuitability for the intended use. In this article, we present the first surrogate deep learning model based on generative adversarial network that predicts thermal residual stress and melt pool characteristics on the cross-section of laser-irradiated carbon steel. The model was trained using simulation data obtained from the unified momentum equation approach-based numerical model. Our model was designed to predict various aspects of the laser melting process only from three scalar inputs of laser power, beam diameter, and irradiation time at two different points in time (immediately after the laser heating is finished and when the steel is completely cooled down) and successfully predicted the distributions of 12 variables, including residual stress and temperature. Prediction results were not only qualitatively very similar to their ground truth, but also quantitatively accurate with the R2 values ranging from 0.975 to 0.999 and the mean absolute errors between 0.0015 and 0.0126. We demonstrated that complex experiments can be quickly performed using the surrogate model because a single deep learning prediction took only 0.13 s, compared to the 2.5–8 h numerical simulation times.

Abstract PS07-02: Benign breast disease and microcalcifications in percutaneous biopsies and breast cancer risk

Abstract Objectives: Mammographic microcalcifications (calcs) are a sentinel marker of in-situ and invasive breast cancer (BC) but most calcs reflect benign breast disease (BBD). BBD comprises a diverse set of lesions, variably associated with increased BC risk. While calc frequency data for the modern era are limited, historically, calcs were reported in 11% of BBD biopsies. Some studies suggest that calcs in BBD impact BC risks; however, the frequency and BC risks associated with calcs during the percutaneous biopsy era have not been clearly established. Accordingly, we assessed the association of calcs with specific BBD lesions and BC risk in a large contemporaneous BBD cohort including detailed pathology review. Methods: A cohort of 4,819 patients with percutaneously diagnosed BBD from 2002-2013 at Mayo Clinic, were identified. The cohort was followed from 6 months after biopsy until censoring, BC diagnosis, or December 2021. Mean follow-up was 11.3 (SD=5.1) years for BBD patients who remained cancer-free in follow-up compared with 7.4 (SD=4.3) years for cases. Histology of all biopsies were reviewed microscopically to record pathologic findings. A natural language processing algorithm (NLPA) was developed and utilized to identify the presence of calcs based on pathology reports. NLPA results were compared to 300 matched biopsy reports, taken as ground truth. We performed age-adjusted Cox proportional hazards regression to assess hazard ratios (HRs) with 95% confidence intervals (CIs) for specific BBD lesions, stratified by the presence (calc+) or absence of calcs (calcs-), with estimation of an interaction term to test for differences in risks by calc strata. Standardized incidence ratios (SIRs) estimated risks for calc+ and calc- BBD lesions versus matched strata for Iowa SEER population-based BC incidence rates. Results: The NLPA achieved sensitivity=0.993 and specificity=0.986. Overall, 42.8% of BBD biopsies were calc+, including 49.1% of biopsies preceding BC and 42.4% of biopsies not preceding BC (Chi-square p=0.02). Overall, risks of BC among women with BBD calc+ biopsies did not differ significantly from those of women with calc- BBD: HR=1.13 (95% CI: 0.90,1.41), p=0.290. SIRs were also similar for the two groups; women with BBD calc+ biopsies had an SIR =2.00 (95% CI: 1.72, 2.33) and BBD calc- biopsies yielded an SIR= 1.91 (95% CI: 1.64,2.22) (Table 1). HRs were similar for specific calc+ and calc- BBD lesions, with progressively higher risks for non-proliferative disease (NP), proliferative disease without atypia (PDWA) and atypical hyperplasia (AH) (pinteraction=0.55). Interaction terms were not significant for any specific lesion, apart from a marginally significant result for dilated ducts, wherein risks were marginally higher for calc- lesions (pInteraction =0.048), perhaps representing a chance finding. SIRs for calc+ versus calc- BBD lesions were also similar. Overall, 77.3% of AH was calc+; for AH calc+, SIR=3.76 (95%CI: 2.74-5.17) and for AH calc-, SIR=4.41 (95%CI: 2.56-7.59). Conclusion: In this recent, larger percutaneous BBD cohort, calcs were present in 42.8% of all specimens, representing a marked increase compared with historical BBD data, particularly those dominated by surgical biopsy for diagnosis and preceding implementation of full-field digital mammography. In this study, calcs in percutaneously diagnosed BBD were not associated with increased BC risk. Table 1: Comparison of standard incidence ratio of subsequent breast cancer with and without microcalcifications in benign breast disease by histologic impression (NP = non-proliferative disease, PDWA = proliferative disease without atypia, and AH = atypical hyperplasia). (1) compared to Iowa SEER population Citation Format: Sarah Schrup, Robert Vierkant, Stacey Winham, Celine Vachon, Derek Radisky, Jodi Carter, Laura Pacheco-Spann, Matthew Jensen, Amy Degnim, Mark Sherman. Benign breast disease and microcalcifications in percutaneous biopsies and breast cancer risk [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PS07-02.