Small sized centroblasts as poor prognostic factor in follicular lymphoma - Based on artificial intelligence analysis.

Histological assessment of centroblasts is an important evaluation in the diagnosis of follicular lymphoma, but there is substantial observer variation in assessment among hematopathologists. We aimed to perform quantitative morphological analysis of centroblasts in follicular lymphoma using new artificial intelligence technology in relation to the clinical prognosis. Hematoxylin and eosin slides of lesions were prepared from 36 cases of follicular lymphoma before initial chemotherapy. Cases were classified into three groups by clinical course after initial treatment. The 'excellent prognosis' group were without recurrence or progression of follicular lymphoma within 60 months, the 'poor prognosis' group were those that had relapse, exacerbation, or who died due to the follicular lymphoma within 60 months, and the 'indeterminate prognosis' group were those without recurrence or progression but before the passage of 60 months. We created whole slide images and image patches of hematoxylin and eosin sections for all cases. We designed an object detection model specialized for centroblasts by fine-tuning YOLOv5 and segmented all centroblasts in whole slide images. The morphological characteristics of centroblasts in relation to the clinical prognosis of follicular lymphoma were analyzed. Centroblasts in follicular lymphoma of the poor prognosis group were significantly smaller in nuclear size than those in follicular lymphoma of the excellent prognosis group in the following points: median of nuclear area (p = 0.013), long length (p = 0.042), short length (p = 0.007), nuclear area of top 10 % cells (p = 0.024) and short length of top 10 % cells (p = 0.020). Cases with a mean nuclear area of <55 µm2 had poorer event-free survival than those with a mean nuclear area of ≥55 µm2 (p < 0.0123). AI methodology is suggested to be able to surpass pathologist's observation in capturing morphological features. Small-sized centroblasts will likely become a new prognostic factor of follicular lymphoma.

New generative methods for single-cell transcriptome data in bulk RNA sequence deconvolution.

Numerous methods for bulk RNA sequence deconvolution have been developed to identify cellular targets of diseases by understanding the composition of cell types in disease-related tissues. However, issues of heterogeneity in gene expression between subjects and the shortage of reference single-cell RNA sequence data remain to achieve accurate bulk deconvolution. In our study, we investigated whether a new data generative method named sc-CMGAN and benchmarking generative methods (Copula, CTGAN and TVAE) could solve these issues and improve the bulk deconvolutions. We also evaluated the robustness of sc-CMGAN using three deconvolution methods and four public datasets. In almost all conditions, the generative methods contributed to improved deconvolution. Notably, sc-CMGAN outperformed the benchmarking methods and demonstrated higher robustness. This study is the first to examine the impact of data augmentation on bulk deconvolution. The new generative method, sc-CMGAN, is expected to become one of the powerful tools for the preprocessing of bulk deconvolution.

Artificial Intelligence Detected the Relationship Between Nuclear Morphological Features and Molecular Abnormalities of Papillary Thyroid Carcinoma.

Papillary thyroid carcinoma (PTC) is the most common type of thyroid carcinoma and has characteristic nuclear features. Genetic abnormalities of PTC affect recent molecular target therapeutic strategy towards RET-altered cases, and they affect clinical prognosis and progression. However, there has been insufficient objective analysis of the correlation between genetic abnormalities and nuclear features. Using our newly developed methods, we studied the correlation between nuclear morphology and molecular abnormalities of PTC with the aim of predicting genetic abnormalities of PTC. We studied 72 cases of PTC and performed genetic analysis to detect BRAF p.V600E mutation and RET fusions. Nuclear features of PTC, such as nuclear grooves, pseudo-nuclear inclusions, and glassy nuclei, were also automatically detected by deep learning models. After analyzing the correlation between genetic abnormalities and nuclear features of PTC, logistic regression models could be used to predict gene abnormalities. Nuclear features were accurately detected with over 0.90 of AUCs in every class. The ratio of glassy nuclei to nuclear groove and the ratio of pseudo-nuclear inclusion to glassy nuclei were significantly higher in cases that were positive for RET fusions (p = 0.027, p = 0.043, respectively) than in cases that were negative for RET fusions. RET fusions were significantly predicted by glassy nuclei/nuclear grooves, pseudo-nuclear inclusions/glassy nuclei, and age (p = 0.023). Our deep learning models could accurately detect nuclear features. Genetic abnormalities had a correlation with nuclear features of PTC. Furthermore, our artificial intelligence model could significantly predict RET fusions of classic PTC.

Use of Artificial Intelligence for the Interpretable Prediction of the Pathologic Diagnosis and Molecular Abnormalities of Flat Urothelial Lesions.

Flat urothelial lesions are important because of their potential for carcinogenesis and development into invasive urothelial carcinomas. However, it is difficult for pathologists to detect early flat urothelial changes and accurately diagnose flat urothelial lesions. To predict the pathologic diagnosis and molecular abnormalities of flat urothelial lesions from pathologic images, artificial intelligence with an interpretable method was used. Next-generation sequencing on 110 hematoxylin and eosin-stained slides of normal urothelium and flat urothelial lesions, including atypical urothelium, dysplasia, and carcinoma in situ, detected 17 types of molecular abnormalities. To generate an interpretable prediction, a new method for segmenting urothelium and a new pathologic criteria-based artificial intelligence (PCB-AI) model was developed. κ Statistics and accuracy measurements were used to evaluate the ability of the model to predict the pathologic diagnosis. The likelihood ratio test was performed to evaluate the logistic regression models for predicting molecular abnormalities. The diagnostic prediction of the PCB-AI model was almost in perfect agreement with the pathologists' diagnoses (weighted κ = 0.98). PCB-AI significantly predicted some molecular abnormalities in an interpretable manner, including abnormalities of TP53 (P = 0.02), RB1 (P = 0.04), and ERCC2 (P = 0.04). Thus, this study developed a new method of obtaining accurate urothelial segmentation, interpretable prediction of pathologic diagnosis, and interpretable prediction of molecular abnormalities.

Pathologic Image Classification of Flat Urothelial Lesions Using Pathologic Criteria-Based Deep Learning.

OBJECTIVES: Pathologic diagnosis of flat urothelial lesions is subject to high interobserver variability. We expected that deep learning could improve the accuracy and consistency of such pathologic diagnosis, although the learning process is a black box. We therefore propose a new approach for pathologic image classification incorporating the diagnostic process of the pathologist into a deep learning method. METHODS: A total of 267 H&E-stained slides of normal urothelium and urothelial lesions from 127 cases were examined. Six independent convolutional neural networks were trained to classify pathologic images according to six pathologic criteria. We then used these networks in the main training for the final diagnosis. RESULTS: Compared with conventional manual analysis, our method significantly improved the classification accuracy of images of flat urothelial lesions. The automated classification showed almost perfect agreement (weighted κ = 0.98) with the consensus reading. In addition, our approach provides the advantages of reliable diagnosis corresponding to histologic interpretation. CONCLUSIONS: We used deep learning to establish an automated subtype classifier for flat urothelial lesions that successfully combines traditional morphologic approaches and complex deep learning to achieve a learning mechanism that seems plausible to the pathologist.