Andy N. D. Nguyen

Recently, vision transformers were shown to be capable of outperforming convolutional neural networks when pretrained on sufficiently large datasets. Vision transformer models show good accuracy on large scale datasets, with features of multi-modal training. Due to their promising feature detection, we aim to explore vision transformer models for diagnosis of anaplastic large cell lymphoma versus classical Hodgkin lymphoma using pathology whole slide images of HE slides. We compared the classification performance of the vision transformer to our previously designed convolutional neural network on the same dataset. The dataset includes whole slide images of HE slides for 20 cases, including 10 cases in each diagnostic category. From each whole slide image, 60 image patches having size of 100 by 100 pixels and at magnification of 20 were obtained to yield 1200 image patches, from which 90 percent were used for training, 9 percent for validation, and 10 percent for testing. The test results from the convolutional neural network model had previously shown an excellent diagnostic accuracy of 100 percent. The test results from the vision transformer model also showed a comparable accuracy at 100 percent. To the best of the authors' knowledge, this is the first direct comparison of predictive performance between a vision transformer model and a convolutional neural network model using the same dataset of lymphoma. Overall, convolutional neural network has a more mature architecture than vision transformer and is usually the best choice when large scale pretraining is not an available option. Nevertheless, our current study shows comparable and excellent accuracy of vision transformer compared to that of convolutional neural network even with a relatively small dataset of anaplastic large cell lymphoma and classical Hodgkin lymphoma.

Deep learning has been shown to be useful in detecting breast cancer metastases by analyzing whole slide images (WSI) of sentinel lymph nodes; however, it requires extensive analysis of all the lymph node slides. Our deep learning study attempts to provide a rapid screen for metastasis by analyzing only a small set of image patches to detect changes in tumor environment. We designed a convolutional neural network to build a diagnostic model for metastasis detection. We obtained WSIs of Hematoxylin and Eosin-stained slides from 34 cases with equal distribution in positive/negative categories. Two WSIs were selected from each case for a total of 69 WSIs. From each WSI, 40 image patches (100x100 pixels) were obtained to yield 2720 image patches, from which 2160 (79%) were used for training, 240 (9%) for validation, and 320 (12%) for testing. Interobserver variation was also examined among 3 users. The test results showed excellent diagnostic results: accuracy (91.15%), sensitivity (77.92%), and specificity (92.09%). No significant variation in results was observed among the 3 observers. This preliminary study provided a proof of concept for conducting a rapid screen for metastasis rather than an exhaustive search for tumors in all fields of all sentinel lymph nodes.

Deep learning has been shown to be useful to detect breast cancer metastases by analyzing whole slide images of sentinel lymph nodes. However, it requires extensive scanning and analysis of all the lymph nodes slides for each case. Our deep learning study focuses on breast cancer screening with only a small set of image patches from any sentinel lymph node, positive or negative for metastasis, to detect changes in tumor environment and not in the tumor itself. We design a convolutional neural network in the Python language to build a diagnostic model for this purpose. The excellent results from this preliminary study provided a proof of concept for incorporating automated metastatic screen into the digital pathology workflow to augment the pathologists' productivity. Our approach is unique since it provides a very rapid screen rather than an exhaustive search for tumor in all fields of all sentinel lymph nodes.

Primary effusion lymphoma (PEL) is a rare type of large B-cell lymphoma associated with human herpesvirus 8 (HHV8) infection. Patients with PEL usually present with an effusion, but occasionally with an extracavitary mass. In this study, we reported a cohort of 70 patients with PEL: 67 men and 3 women with a median age of 46 years (range 26–91). Of these, 56 (80%) patients had human immunodeficiency virus (HIV) infection, eight were HIV-negative, and six had unknown HIV status. Nineteen (27%) patients had Kaposi sarcoma. Thirty-five (50%) patients presented with effusion only, 27 (39%) had an extracavitary mass or masses only, and eight (11%) had both effusion and extracavitary disease. The lymphoma cells showed plasmablastic, immunoblastic, or anaplastic morphology. All 70 (100%) cases were positive for HHV8. Compared with effusion-only PEL, patients with extracavitary-only PEL were younger (median age, 42 vs. 52 years, p = 0.001), more likely to be HIV-positive (88.9% vs. 68.6%, p = 0.06) and EBV-positive (76.9% vs. 51.9%, p = 0.06), and less often positive for CD45 (69.2% vs. 96.2%, p = 0.01), EMA (26.7% vs. 100%, p = 0.0005), and CD30 (60% vs. 81.5%, p = 0.09). Of 52 (50%) patients with clinical follow-up, 26 died after a median follow-up time of 40.0 months (range 0–96), and the median overall survival was 42.5 months. The median OS for patients with effusion-only and with extracavitary-only PEL were 30.0 and 37.9 months, respectively (p = 0.34), and patients with extracavitary-only PEL had a lower mortality rate at the time of last follow-up (35% vs. 61.5%, p = 0.07). The median OS for HIV-positive and HIV-negative patients were 42.5 and 6.8 months, respectively (p = 0.57), and they had a similar mortality rate of 50% at last follow-up. In conclusion, patients presenting with effusion-only versus extracavitary-only disease are associated with different clinicopathologic features. PEL is an aggressive lymphoma with a poor prognosis, regardless of extracavitary presentation or HIV status.

Objective: We explored how Deep Learning can be utilized to predict the prognosis of acute myeloid leukemia. Methods: Out of The Cancer Genome Atlas database, 94 acute myeloid leukemia cases were used in this study. Input data included age, 10 most common cytogenetic and 23 most common mutation results; output was the prognosis (diagnosis to death). In our Deep Learning network, autoencoders were stacked to form a hierarchical Deep Learning model from which raw data were compressed and organized, and high-level features were extracted. The network was written in R language and was designed to predict the prognosis of acute myeloid leukemia for a given case (diagnosis to death of either more or less than 730 days). Results: The Deep Learning network achieved an excellent accuracy of 83% in predicting prognosis. Conclusion: As a proof-of-concept study, our preliminary results demonstrated a practical application of Deep Learning in the future practice of prognostic prediction using next-generation sequencing data.

Recent studies have shown promising results in using Deep Learning to detect malignancy in whole slide imaging, however, they were limited to just predicting a positive or negative finding for a specific neoplasm. We attempted to use Deep Learning with a convolutional neural network (CNN) algorithm to build a lymphoma diagnostic model for four diagnostic categories: (1) benign lymph node, (2) diffuse large B-cell lymphoma, (3) Burkitt lymphoma, and (4) small lymphocytic lymphoma. Our software was written in Python language. We obtained digital whole-slide images of Hematoxylin and Eosin stained slides of 128 cases including 32 cases for each diagnostic category. Four sets of 5 representative images, 40x40 pixels in dimension, were taken for each case. A total of 2,560 images were obtained from which 1,856 were used for training, 464 for validation, and 240 for testing. For each test set of 5 images, the predicted diagnosis was combined from the prediction of five images. The test results showed excellent diagnostic accuracy at 95% for image-by-image prediction and at 100% for set-by-set prediction. This preliminary study provided a proof of concept for incorporating automated lymphoma diagnostic screen into future pathology work-flow to augment the pathologists' productivity.

Deep Learning can significantly benefit cancer proteomics and genomics. In this study, we attempted to determine a set of critical proteins that were associated with the FLT3-ITD mutation in newly-diagnosed acute myeloid leukemia patients. A Deep Learning network consisting of autoencoders formed a hierarchical model from which high-level features were extracted without labeled training data. Dimensional reduction reduced the number of critical proteins from 231 to 20. Deep Learning found an excellent correlation between FLT3-ITD mutation with the levels of these 20 critical proteins (accuracy 97%, sensitivity 90%, and specificity 100%). Our Deep Learning network could hone in on 20 proteins with the strongest association with FLT3-ITD. The results of this study allow for a novel approach to determine critical protein pathways in the FLT3-ITD mutation, and provide proof-of-concept for an accurate approach to model big data in cancer proteomics and genomics.

Recent studies have shown promising results in using Deep Learning to detect malignancy in whole slide imaging. However, they were limited to just predicting positive or negative finding for a specific neoplasm. We attempted to use Deep Learning with a convolutional neural network algorithm to build a lymphoma diagnostic model for four diagnostic categories: benign lymph node, diffuse large B cell lymphoma, Burkitt lymphoma, and small lymphocytic lymphoma. Our software was written in Python language. We obtained digital whole slide images of Hematoxylin and Eosin stained slides of 128 cases including 32 cases for each diagnostic category. Four sets of 5 representative images, 40x40 pixels in dimension, were taken for each case. A total of 2,560 images were obtained from which 1,856 were used for training, 464 for validation and 240 for testing. For each test set of 5 images, the predicted diagnosis was combined from prediction of 5 images. The test results showed excellent diagnostic accuracy at 95% for image-by-image prediction and at 10% for set-by-set prediction. This preliminary study provided a proof of concept for incorporating automated lymphoma diagnostic screen into future pathology workflow to augment the pathologists' productivity.

We explore how Deep Learning (DL) can be utilized to predict prognosis of acute myeloid leukemia (AML). Out of TCGA (The Cancer Genome Atlas) database, 94 AML cases are used in this study. Input data include age, 10 common cytogenetic and 23 most common mutation results; output is the prognosis (diagnosis to death, DTD). In our DL network, autoencoders are stacked to form a hierarchical DL model from which raw data are compressed and organized and high-level features are extracted. The network is written in R language and is designed to predict prognosis of AML for a given case (DTD of more than or less than 730 days). The DL network achieves an excellent accuracy of 83% in predicting prognosis. As a proof-of-concept study, our preliminary results demonstrate a practical application of DL in future practice of prognostic prediction using next-gen sequencing (NGS) data.