Subhajit Chakrabarty1, Karen Stokes2, Junaid Ansari3, Caleb Stewart2, Steven Conrad2, Shashank Shekhar4, Candace Howard-Claudio4, David P. Gordy4 and Kunal Bhatia4
1Computer Science, Louisiana State University (LSU) Shreveport, Shereveport, LA, 2LSU Health Shreveport, Shreveport, LA, 3Johns Hopkins Medicine, Baltimore, MD, 4University of Mississippi Medical Center, Jackson, MS

Introduction/Background. Deep Learning may provide automated tools for fast diagnosis. Our primary context was studying brain MRIs in patients with ischemic stroke and COVID-19, using state-of-the-art Deep Learning tools.

Hypothesis/Goal of Study. The first aim was to perform automatic lesion segmentation in our context. The second aim was to classify strokes. The third aim was to classify COVID vs. non-COVID, in our context.

Methods and Results. Our dataset was the ATLAS 2.0 dataset that has manually segmented lesion masks, as well as our own-collected data. We used 600 T1-weighted MRI scans for patients with stroke (for COVID and non-COVID). For the first task, we derived our model from U-Net architecture. For the second and third tasks, we derived our model from Vision Transformer architecture.

Discussion/Conclusions. Results indicated good performance in both segmentation and classification tasks. We also explored the association with features such as free-floating thrombi, intraventricular hemorrhage, thalamic venous stroke, and multifocal ischemia.

Citation/Acknowledgements. Research reported in this presentation was supported by an Institutional Development Award (IDeA) from the National Institute of General Medical Sciences of the National Institutes of Health, through LBRN Start-up Funds 2022-24 awarded to Dr. Subhajit Chakrabarty, and a matching grant provided through LSU Shreveport.

Sarah Floyd1, Ahmed Alameldin2, Hudson Smith3, Sydney Lindros1, Kyle Jeray5, and Jihad Obeid4
1Department of Public Health Sciences, Clemson University, Clemson SC, 2College of Computing, Clemson University, Clemson SC, 3Director of Applied Machine Learning, Clemson University, Clemson SC, 4Department of Public Health Sciences, Medical University of South Carolina, Charleston SC, 5Department of Orthopedic Surgery, Prisma Health, Greenville SC

Introduction/Background. Outcome data is not routinely collected for patients with proximal humerus fractures (PHF), which limits our ability to conduct comparative treatment studies. The automatic classification of clinical notes has emerged as a tool to extract meaningful insights from unstructured text data.

Hypothesis/Goal of Study. This study aims to develop a new orthopaedic treatment success measure derived from unstructured clinical notes recorded in electronic health records (EHR).

Methods and Results. A sample of patients with an ICD-10 confirmed PHF were identified between 2019 and 2022 at one health system. For each patient, we identified the last orthopaedic encounter each patient had within 365 days of their index fracture visit. The complete clinical note was reviewed and labeled by four orthopaedic residents. The annotated clinical notes served as the gold standard for developing and validating the text classifiers. Several machine learning models were tested, and the area under the receiver operating curve (AUC) was used to fine tune and assess the performance of each model. A 10-fold cross-validation was conducted to provide an unbiased estimation of the model prediction performance. Results A sample of 868 unique patients with confirmed PHF were identified and composed the study sample. Of the 868, 465 (53.6%) were labeled as treatment success, and the remaining 46.4% were labeled as treatment failure due to ongoing symptoms or complications from treatment. Interrater agreement for classification of clinical notes between residents was moderate (pairwise percent agreement = 75.31%, Fleiss’ κ = 0.49 (0.30-0.68)). The linear Support Vector Classification model had the best performance with an AUC of 0.90.

Discussion/Conclusions. Our newly developed classifier can identify treatment success using unstructured EHR clinical notes. This outcome measure will facilitate the development of clinical decision support tools to provide personalized care recommendations for patients with PHF.

Grant/Funding Support. Agency for Healthcare Research and Quality 1R03HS029060-01 and P20 GM121342

Acknowledgements. We thank the four orthopaedic residents responsible for the labeling of the orthopaedic notes, Benjamin Judkins, Zach Reynolds, Claire Krohn and Jahan Threeths.

Alireza Bagheri Rajeoni1, Breanna Pederson2, Susan M. Lessner2, and Homayoun Valafar1
1Department of Computer Science and Engineering, University of South Carolina, Columbia, SC and 2University of South Carolina School of Medicine
Columbia, University of South Carolina, Columbia, SC

Introduction/Background. The study aims to address the time-consuming and exhaustive task of manually examining medical images, such as computed tomographic angiograms (CTAs), to analyze the vascular system in patients undergoing surgery for peripheral arterial disease (PAD). We propose a deep learning model tailored to segment vascultare in CTA images starting from the descending thoracic aorta to the knees.The model incorporates advanced deep learning techniques to achieve precise segmentation of these regions, expediting the vascular analysis process and potentially aiding in better diagnostic and treatment outcomes.

Hypothesis/Goal of Study. Numerous chronic diseases, including atherosclerosis and aneurysms are rooted in abnormal changes within the human vascular system. However, the manual examination of medical images, such as computed tomographic angiograms (CTAs), for analyzing the vascular system is a time-consuming and exhaustive task. To tackle this challenge, we propose a deep learning model specifically designed to segment the vascular system in CTA images of patients who undergo surgery for peripheral arterial disease (PAD). Our research focuses on accurately predicting two regions: (1) from the descending thoracic aorta to the iliac bifurcation, and (2) from the descending thoracic aorta to the knees in CTA images, utilizing advanced deep learning techniques.

Methods and Results. Using the dataset of 11 patients collected at Prisma health 1, we utilized a deep learning algorithm to segment the vascular system. Our model architecture follows an encoder-decoder structure similar to U-net 2, incorporating skip connections from the encoder to decoder. Additionally, a Transformer 3 is utilized in the bridge between these two main blocks as illustrated in Figure. The inputs to the model were images with dimensions of 512x512 and three channels, and the output was a mask with dimensions of 512x512 and one channel. We trained the model using a batch size of 40, for 400 epochs, with a learning rate of 1e-3 During the training and validation process, we achieved average IOU accuracies of 97.3% and 95% for segmenting the vascular system from the descending thoracic aorta to the knees, respectively, in the cross-validation. In the testing dataset, we obtained average Dice accuracies of 93.3% and 83.4% for (1) segmenting the vascular system from the descending thoracic aorta to the iliac bifurcation and (2) from the descending thoracic aorta to the knees, respectively.

Discussion/Conclusions. Accurately capturing and examining the vascular system enables the identification of various pathological conditions like aneurysms and vascular calcification. Moving forward, our primary objective is to improve the accuracy of segmenting and precisely measuring calcification within the vascular system. This progress will significantly enhance diagnostic precision, facilitate proactive treatment, and ultimately lead to better outcomes for patients in the field of vascular health.

Citation/Acknowledgements. This work was funded by NIH grant number P20 RR-016461 to Dr. Valafar and HL145064-01 to Dr. Lessner. This work was also partially supported by the National Science Foundation EPSCoR Program under NSF Award # OIA-2242812. 1. Zhao, L., Odigwe, B., Lessner, S., Clair, D., Mussa, F., & Valafar, H. (2019). Automated Analysis of Femoral Artery Calcification Using Machine Learning Techniques. 2019 International Conference on Computational Science and Computational Intelligence (CSCI), 584–589. https://doi.org/10.1109/CSCI49370.2019.00110 2. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. In N. Navab, J. Hornegger, W. M. Wells, & A. F. Frangi (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015 (pp. 234–241). Springer International Publishing. https://doi.org/10.1007/978-3-319-24574-4_28 3. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., & Houlsby, N. (2021). An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (arXiv:2010.11929). arXiv. https://doi.org/10.48550/arXiv.2010.11929