More Like this

1. The Temple University Digital Pathology Corpus: The Breast Tissue Subset

   https://doi.org/10.1109/SPMB52430.2021.9672275  

   Wevodau, Z. ; Doshna, B. ; Jhala, N. ; Akhtar, I. ; Obeid, I. ; Picone, J. ( December 2021 , Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB))

   Obeid, I. (Ed.)

   The Neural Engineering Data Consortium (NEDC) is developing the Temple University Digital Pathology Corpus (TUDP), an open source database of high-resolution images from scanned pathology samples [1], as part of its National Science Foundation-funded Major Research Instrumentation grant titled “MRI: High Performance Digital Pathology Using Big Data and Machine Learning” [2]. The long-term goal of this project is to release one million images. We have currently scanned over 100,000 images and are in the process of annotating breast tissue data for our first official corpus release, v1.0.0. This release contains 3,505 annotated images of breast tissue including 74 patients with cancerous diagnoses (out of a total of 296 patients). In this poster, we will present an analysis of this corpus and discuss the challenges we have faced in efficiently producing high quality annotations of breast tissue. It is well known that state of the art algorithms in machine learning require vast amounts of data. Fields such as speech recognition [3], image recognition [4] and text processing [5] are able to deliver impressive performance with complex deep learning models because they have developed large corpora to support training of extremely high-dimensional models (e.g., billions of parameters). Other fields that do not have access to such data resources must rely on techniques in which existing models can be adapted to new datasets [6]. A preliminary version of this breast corpus release was tested in a pilot study using a baseline machine learning system, ResNet18 [7], that leverages several open-source Python tools. The pilot corpus was divided into three sets: train, development, and evaluation. Portions of these slides were manually annotated [1] using the nine labels in Table 1 [8] to identify five to ten examples of pathological features on each slide. Not every pathological feature is annotated, meaning excluded areas can include focuses particular to these labels that are not used for training. A summary of the number of patches within each label is given in Table 2. To maintain a balanced training set, 1,000 patches of each label were used to train the machine learning model. Throughout all sets, only annotated patches were involved in model development. The performance of this model in identifying all the patches in the evaluation set can be seen in the confusion matrix of classification accuracy in Table 3. The highest performing labels were background, 97% correct identification, and artifact, 76% correct identification. A correlation exists between labels with more than 6,000 development patches and accurate performance on the evaluation set. Additionally, these results indicated a need to further refine the annotation of invasive ductal carcinoma (“indc”), inflammation (“infl”), nonneoplastic features (“nneo”), normal (“norm”) and suspicious (“susp”). This pilot experiment motivated changes to the corpus that will be discussed in detail in this poster presentation. To increase the accuracy of the machine learning model, we modified how we addressed underperforming labels. One common source of error arose with how non-background labels were converted into patches. Large areas of background within other labels were isolated within a patch resulting in connective tissue misrepresenting a non-background label. In response, the annotation overlay margins were revised to exclude benign connective tissue in non-background labels. Corresponding patient reports and supporting immunohistochemical stains further guided annotation reviews. The microscopic diagnoses given by the primary pathologist in these reports detail the pathological findings within each tissue site, but not within each specific slide. The microscopic diagnoses informed revisions specifically targeting annotated regions classified as cancerous, ensuring that the labels “indc” and “dcis” were used only in situations where a micropathologist diagnosed it as such. Further differentiation of cancerous and precancerous labels, as well as the location of their focus on a slide, could be accomplished with supplemental immunohistochemically (IHC) stained slides. When distinguishing whether a focus is a nonneoplastic feature versus a cancerous growth, pathologists employ antigen targeting stains to the tissue in question to confirm the diagnosis. For example, a nonneoplastic feature of usual ductal hyperplasia will display diffuse staining for cytokeratin 5 (CK5) and no diffuse staining for estrogen receptor (ER), while a cancerous growth of ductal carcinoma in situ will have negative or focally positive staining for CK5 and diffuse staining for ER [9]. Many tissue samples contain cancerous and non-cancerous features with morphological overlaps that cause variability between annotators. The informative fields IHC slides provide could play an integral role in machine model pathology diagnostics. Following the revisions made on all the annotations, a second experiment was run using ResNet18. Compared to the pilot study, an increase of model prediction accuracy was seen for the labels indc, infl, nneo, norm, and null. This increase is correlated with an increase in annotated area and annotation accuracy. Model performance in identifying the suspicious label decreased by 25% due to the decrease of 57% in the total annotated area described by this label. A summary of the model performance is given in Table 4, which shows the new prediction accuracy and the absolute change in error rate compared to Table 3. The breast tissue subset we are developing includes 3,505 annotated breast pathology slides from 296 patients. The average size of a scanned SVS file is 363 MB. The annotations are stored in an XML format. A CSV version of the annotation file is also available which provides a flat, or simple, annotation that is easy for machine learning researchers to access and interface to their systems. Each patient is identified by an anonymized medical reference number. Within each patient’s directory, one or more sessions are identified, also anonymized to the first of the month in which the sample was taken. These sessions are broken into groupings of tissue taken on that date (in this case, breast tissue). A deidentified patient report stored as a flat text file is also available. Within these slides there are a total of 16,971 total annotated regions with an average of 4.84 annotations per slide. Among those annotations, 8,035 are non-cancerous (normal, background, null, and artifact,) 6,222 are carcinogenic signs (inflammation, nonneoplastic and suspicious,) and 2,714 are cancerous labels (ductal carcinoma in situ and invasive ductal carcinoma in situ.) The individual patients are split up into three sets: train, development, and evaluation. Of the 74 cancerous patients, 20 were allotted for both the development and evaluation sets, while the remain 34 were allotted for train. The remaining 222 patients were split up to preserve the overall distribution of labels within the corpus. This was done in hope of creating control sets for comparable studies. Overall, the development and evaluation sets each have 80 patients, while the training set has 136 patients. In a related component of this project, slides from the Fox Chase Cancer Center (FCCC) Biosample Repository (https://www.foxchase.org/research/facilities/genetic-research-facilities/biosample-repository -facility) are being digitized in addition to slides provided by Temple University Hospital. This data includes 18 different types of tissue including approximately 38.5% urinary tissue and 16.5% gynecological tissue. These slides and the metadata provided with them are already anonymized and include diagnoses in a spreadsheet with sample and patient ID. We plan to release over 13,000 unannotated slides from the FCCC Corpus simultaneously with v1.0.0 of TUDP. Details of this release will also be discussed in this poster. Few digitally annotated databases of pathology samples like TUDP exist due to the extensive data collection and processing required. The breast corpus subset should be released by November 2021. By December 2021 we should also release the unannotated FCCC data. We are currently annotating urinary tract data as well. We expect to release about 5,600 processed TUH slides in this subset. We have an additional 53,000 unprocessed TUH slides digitized. Corpora of this size will stimulate the development of a new generation of deep learning technology. In clinical settings where resources are limited, an assistive diagnoses model could support pathologists’ workload and even help prioritize suspected cancerous cases. ACKNOWLEDGMENTS This material is supported by the National Science Foundation under grants nos. CNS-1726188 and 1925494. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. REFERENCES [1] N. Shawki et al., “The Temple University Digital Pathology Corpus,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York City, New York, USA: Springer, 2020, pp. 67 104. https://www.springer.com/gp/book/9783030368432. [2] J. Picone, T. Farkas, I. Obeid, and Y. Persidsky, “MRI: High Performance Digital Pathology Using Big Data and Machine Learning.” Major Research Instrumentation (MRI), Division of Computer and Network Systems, Award No. 1726188, January 1, 2018 – December 31, 2021. https://www. isip.piconepress.com/projects/nsf_dpath/. [3] A. Gulati et al., “Conformer: Convolution-augmented Transformer for Speech Recognition,” in Proceedings of the Annual Conference of the International Speech Communication Association (INTERSPEECH), 2020, pp. 5036-5040. https://doi.org/10.21437/interspeech.2020-3015. [4] C.-J. Wu et al., “Machine Learning at Facebook: Understanding Inference at the Edge,” in Proceedings of the IEEE International Symposium on High Performance Computer Architecture (HPCA), 2019, pp. 331–344. https://ieeexplore.ieee.org/document/8675201. [5] I. Caswell and B. Liang, “Recent Advances in Google Translate,” Google AI Blog: The latest from Google Research, 2020. [Online]. Available: https://ai.googleblog.com/2020/06/recent-advances-in-google-translate.html. [Accessed: 01-Aug-2021]. [6] V. Khalkhali, N. Shawki, V. Shah, M. Golmohammadi, I. Obeid, and J. Picone, “Low Latency Real-Time Seizure Detection Using Transfer Deep Learning,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2021, pp. 1 7. https://www.isip. piconepress.com/publications/conference_proceedings/2021/ieee_spmb/eeg_transfer_learning/. [7] J. Picone, T. Farkas, I. Obeid, and Y. Persidsky, “MRI: High Performance Digital Pathology Using Big Data and Machine Learning,” Philadelphia, Pennsylvania, USA, 2020. https://www.isip.piconepress.com/publications/reports/2020/nsf/mri_dpath/. [8] I. Hunt, S. Husain, J. Simons, I. Obeid, and J. Picone, “Recent Advances in the Temple University Digital Pathology Corpus,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2019, pp. 1–4. https://ieeexplore.ieee.org/document/9037859. [9] A. P. Martinez, C. Cohen, K. Z. Hanley, and X. (Bill) Li, “Estrogen Receptor and Cytokeratin 5 Are Reliable Markers to Separate Usual Ductal Hyperplasia From Atypical Ductal Hyperplasia and Low-Grade Ductal Carcinoma In Situ,” Arch. Pathol. Lab. Med., vol. 140, no. 7, pp. 686–689, Apr. 2016. https://doi.org/10.5858/arpa.2015-0238-OA. 

   more » « less
2. Searching for Anomalies in the ZTF Catalog of Periodic Variable Stars

   https://doi.org/10.3847/1538-4357/ac69d4  

   Chan, Ho-Sang ; Villar, V. Ashley ; Cheung, Siu-Hei ; Ho, Shirley ; O’Grady, Anna J. G. ; Drout, Maria R. ; Renzo, Mathieu ( June 2022 , The Astrophysical Journal)

   Periodic variables illuminate the physical processes of stars throughout their lifetime. Wide-field surveys continue to increase our discovery rates of periodic variable stars. Automated approaches are essential to identify interesting periodic variable stars for multiwavelength and spectroscopic follow-up. Here we present a novel unsupervised machine-learning approach to hunt for anomalous periodic variables using phase-folded light curves presented in the Zwicky Transient Facility Catalogue of Periodic Variable Stars by Chen et al. We use a convolutional variational autoencoder to learn a low-dimensional latent representation, and we search for anomalies within this latent dimension via an isolation forest. We identify anomalies with irregular variability. Most of the top anomalies are likely highly variable red giants or asymptotic giant branch stars concentrated in the Milky Way galactic disk; a fraction of the identified anomalies are more consistent with young stellar objects. Detailed spectroscopic follow-up observations are encouraged to reveal the nature of these anomalies.

    

   more » « less
3. Gravity Spy Volunteer Classifications of LIGO Glitches from Observing Runs O1, O2, O3a, and O3b

   https://doi.org/10.5281/zenodo.5911226  

   Zevin, Michael ; Coughlin, Scott ; Chase, Eve ; Allen, Sara ; Bahaadini, Sara ; Berry, Christopher ; Crowston, Kevin ; Harandi, Mabi ; Jackson, Corey ; Kalogera, Vicky ; et al ( January 2022 , Zenodo)

   This dataset contains machine learning and volunteer classifications from the Gravity Spy project. It includes glitches from observing runs O1, O2, O3a and O3b that received at least one classification from a registered volunteer in the project. It also indicates glitches that are nominally retired from the project using our default set of retirement parameters, which are described below. See more details in the Gravity Spy Methods paper. 

   When a particular subject in a citizen science project (in this case, glitches from the LIGO datastream) is deemed to be classified sufficiently it is "retired" from the project. For the Gravity Spy project, retirement depends on a combination of both volunteer and machine learning classifications, and a number of parameterizations affect how quickly glitches get retired. For this dataset, we use a default set of retirement parameters, the most important of which are: 

   1. A glitches must be classified by at least 2 registered volunteers
   2. Based on both the initial machine learning classification and volunteer classifications, the glitch has more than a 90% probability of residing in a particular class
   3. Each volunteer classification (weighted by that volunteer's confusion matrix) contains a weight equal to the initial machine learning score when determining the final probability

   The choice of these and other parameterization will affect the accuracy of the retired dataset as well as the number of glitches that are retired, and will be explored in detail in an upcoming publication (Zevin et al. in prep). 

   The dataset can be read in using e.g. Pandas: 
   ```
   import pandas as pd
   dataset = pd.read_hdf('retired_fulldata_min2_max50_ret0p9.hdf5', key='image_db')
   ```
   Each row in the dataframe contains information about a particular glitch in the Gravity Spy dataset. 

   Description of series in dataframe

   • ['1080Lines', '1400Ripples', 'Air_Compressor', 'Blip', 'Chirp', 'Extremely_Loud', 'Helix', 'Koi_Fish', 'Light_Modulation', 'Low_Frequency_Burst', 'Low_Frequency_Lines', 'No_Glitch', 'None_of_the_Above', 'Paired_Doves', 'Power_Line', 'Repeating_Blips', 'Scattered_Light', 'Scratchy', 'Tomte', 'Violin_Mode', 'Wandering_Line', 'Whistle']
     • Machine learning scores for each glitch class in the trained model, which for a particular glitch will sum to unity
   • ['ml_confidence', 'ml_label']
     • Highest machine learning confidence score across all classes for a particular glitch, and the class associated with this score
   • ['gravityspy_id', 'id']
     • Unique identified for each glitch on the Zooniverse platform ('gravityspy_id') and in the Gravity Spy project ('id'), which can be used to link a particular glitch to the full Gravity Spy dataset (which contains GPS times among many other descriptors)
   • ['retired']
     • Marks whether the glitch is retired using our default set of retirement parameters (1=retired, 0=not retired)
   • ['Nclassifications']
     • The total number of classifications performed by registered volunteers on this glitch
   • ['final_score', 'final_label']
     • The final score (weighted combination of machine learning and volunteer classifications) and the most probable type of glitch
   • ['tracks']
     • Array of classification weights that were added to each glitch category due to each volunteer's classification

    

   ```
   For machine learning classifications on all glitches in O1, O2, O3a, and O3b, please see Gravity Spy Machine Learning Classifications on Zenodo

   For the most recently uploaded training set used in Gravity Spy machine learning algorithms, please see Gravity Spy Training Set on Zenodo.

   For detailed information on the training set used for the original Gravity Spy machine learning paper, please see Machine learning for Gravity Spy: Glitch classification and dataset on Zenodo. 

    

   more » « less
4. Kinematics of molecular gas in star-forming galaxies with large-scale ionized outflows

   https://doi.org/10.1093/mnras/stac3083  

   Hogarth, L. M. ; Saintonge, A. ; Davis, T. A. ( November 2022 , Monthly Notices of the Royal Astronomical Society)

   We investigate the kinematics of the molecular gas in a sample of seven edge-on (i > 60°) galaxies identified as hosting large-scale outflows of ionized gas, using ALMA CO(1–0) observations at ∼1 kpc resolution. We build on Hogarth et al., where we find that molecular gas is more centrally concentrated in galaxies which host winds than in control objects. We perform full three-dimensional kinematic modelling with multiple combinations of kinematic components, allowing us to infer whether these objects share any similarities in their molecular gas structure. We use modelling to pinpoint the kinematic centre of each galaxy, in order to interpret their minor- and major-axis position velocity diagrams (PVDs). From the PVDs, we find that the bulk of the molecular gas in our galaxies is dynamically cold, tracing the rotation curves predicted by our symmetric, rotation-dominated models, but with minor flux asymmetries. Most notably, we find evidence of radial gas motion in a subset of our objects, which demonstrate a characteristic ‘twisting’ in their minor-axis PVDs generally associated with gas flow along the plane of a galaxy. In our highest S/N object, we include bi-symmetric radial flow in our kinematic model, and find (via the Bayesian Information Criterion) that the presence of radial gas motion is strongly favoured. This may provide one mechanism by which molecular gas and star formation are centrally concentrated, enabling the launch of massive ionized gas winds. However, in the remainder of our sample, we do not observe evidence that gas is being driven radially, once again emphasizing the variety of physical processes that may be powering the outflows in these objects, as originally noted in H21.

    

   more » « less
5. Radio pulsar beam geometry down to the 100-MHz band: 76 additional sources within the Arecibo Sky

   https://doi.org/10.1093/mnras/stac3025  

   Rankin, Joanna ; Wahl, Haley ; Venkataraman, Arun ; Olszanski, Timothy ( October 2022 , Monthly Notices of the Royal Astronomical Society)

   This paper provides analyses of the emission beam structure of 76 ‘B’-named pulsars within the Arecibo sky. Most of these objects are included in both the Gould & Lyne and LOFAR High Band surveys and thus complement our other works treating various parts of these populations. These comprise a further group of mostly well-studied pulsars within the Arecibo sky that we here treat similarly to those in Olszanski et al. – and extend our overall efforts to study all of the pulsars in both surveys. The analyses are based on observations made with the Arecibo Telescope at 327 MHz and 1.4 GHz. Many have been observed at frequencies down to 100 MHz using either LOFAR or the Pushchino Radio Astronomy Observatory as well as a few with the Long Wavelength Array at lower frequencies. This work uses the Arecibo observations as a foundation for interpreting the low frequency profiles and emission-beam geometries. We attempt to build quantitative geometric emission-beam models using the core/double-cone topology, while reviewing the evidence of previous studies and arguments for previous classifications on these sources. These efforts were successful for all but two pulsars, and interesting new subpulse modulation patterns were identified in a number of the objects. We interpret the Arecibo pulsar population in the context of the entire population of ‘B’ pulsars.

    

   more » « less