16 December 2022 in Cancer Cytopathology
Background: Several studies have used artificial intelligence (AI) to analyze cytology images, but AI has yet to be adopted in clinical practice. The objective of this study was to demonstrate the accuracy of AI‐based image analysis for thyroid fine‐needle aspiration cytology (FNAC) and to propose its application in clinical practice.Methods: In total, 148,395 microscopic images of FNAC were obtained from 393 thyroid nodules to train and validate the data, and EfficientNetV2‐L was used as the image‐classification model. The 35 nodules that were classified as atypia of undetermined significance (AUS) were predicted using AI training.Results: The precision‐recall area under the curve (PR AUC) was >0.95, except for poorly differentiated thyroid carcinoma (PR AUC = 0.49) and medullary thyroid carcinoma (PR AUC = 0.91). Poorly differentiated thyroid carcinoma had the lowest recall (35.4%) and was difficult to distinguish from papillary thyroid carcinoma, medullary thyroid carcinoma, and follicular thyroid carcinoma. Follicular adenomas and follicular thyroid carcinomas were distinguished from each other by 86.7% and 93.9% recall, respectively. For two‐dimensional mapping of the data using t‐distributed stochastic neighbor embedding, the lymphomas, follicular adenomas, and anaplastic thyroid carcinomas were divided into three, two, and two groups, respectively. Analysis of the AUS nodules showed 94.7% sensitivity, 14.4% specificity, 56.3% positive predictive value, and 66.7% negative predictive value.Conclusions: The authors developed an AI‐based approach to analyze thyroid FNAC cases encountered in routine practice. This analysis could be useful for the clinical management of AUS and follicular neoplasm nodules (e.g., an online AI platform for thyroid cytology consultations).
9 October 2024 in Journal of the American Society of Nephrology
Key Points: Self-supervised learning extracts meaningful glomerular features without teacher labels. Self-distillation with no labels–pretrained model outperformed conventional supervised learning in disease and clinical parameter classification. Self-distillation with no labels model enabled deep learning on small datasets, reducing annotation efforts. Background: Deep learning has great potential in digital kidney pathology. However, its effectiveness depends heavily on the availability of extensively labeled datasets, which are often limited because of the specialized knowledge and time required for their creation. This limitation hinders the widespread application of deep learning for the analysis of kidney biopsy images. Methods: We applied self-distillation with no labels (DINO), a self-supervised learning method, to a dataset of 10,423 glomerular images obtained from 384 periodic acid–Schiff-stained kidney biopsy slides. Glomerular features extracted from the DINO-pretrained backbone were visualized using principal component analysis. We then performed classification tasks by adding either k-nearest neighbor classifiers or linear head layers to the DINO-pretrained or ImageNet-pretrained backbones. These models were trained on our labeled classification dataset. Performance was evaluated using metrics such as the area under the receiver operating characteristic curve (ROC-AUC). The classification tasks encompassed four disease categories (minimal change disease, mesangial proliferative GN, membranous nephropathy, and diabetic nephropathy) and clinical parameters such as hypertension, proteinuria, and hematuria. Results: Principal component analysis visualization revealed distinct principal components corresponding to different glomerular structures, demonstrating the capability of the DINO-pretrained backbone to capture morphologic features. In disease classification, the DINO-pretrained transferred model (ROC-AUC=0.93) outperformed the ImageNet-pretrained fine-tuned model (ROC-AUC=0.89). When the labeled data were limited, the ImageNet-pretrained fine-tuned model's ROC-AUC dropped to 0.76 (95% confidence interval, 0.72 to 0.80), whereas the DINO-pretrained transferred model maintained superior performance (ROC-AUC, 0.88; 95% confidence interval, 0.86 to 0.90). The DINO-pretrained transferred model also exhibited higher AUCs for the classification of several clinical parameters. External validation using two independent datasets confirmed DINO pretraining's superiority, particularly when labeled data were limited. Conclusions: The application of DINO to unlabeled periodic acid–Schiff-stained glomerular images facilitated the extraction of histologic features that could be effectively used for disease classification.
3 July 2025 in Communications Medicine
Background Polyploidy is frequently observed in cancer cells and is closely associated with chromosomal instability, which can lead to cancer progression. Polyploid cancers are more aggressive than diploid cancers, and polyploidy has been shown to be a prognostic marker for hepatocellular carcinoma (HCC). However, polyploidy is challenging to diagnose. Currently, no clinically implementable methods are available for diagnosing polyploidy in cancer. Methods We established a method for assessing polyploidization in HCC using deep-learning-based artificial intelligence image recognition models to assess hematoxylin and eosin-stained pathological images. Using 44 HCCs whose ploidy status had been determined by chromosome fluorescence in situ hybridization, we evaluated the ability of our constructed deep learning models to detect HCC ploidy. We then tested the models on an independent group of 169 liver cancers and applied them to a publicly available dataset. Results Here we show that our constructed models effectively assess HCC ploidy in a separate cohort and identify a subset with poor prognosis based on the ploidy determinations for 169 HCCs. Our pipeline also identifies HCCs with poor prognosis in the external dataset, with a more significant difference than that for ploidy inferences by genomic analysis. By exploiting the high processing capacity of artificial intelligence, new aspects of polyploid HCC, such as the high prevalence of scirrhous structures, are identified. Conclusions Our findings suggest that ploidy assessment using artificial intelligence-based pathological image recognition can serve as a novel diagnostic tool for personalized medicine.
6 October 2026 in Communications Medicine
Background: Thyroid fine-needle aspiration cytology is widely used to evaluate thyroid nodules before surgery. Recent advances in deep learning have increased the potential for artificial intelligence-assisted cytological diagnosis. However, developing accurate models usually requires large numbers of images labelled by experts, making dataset preparation time-consuming and costly. We investigate whether self-supervised learning can improve thyroid cytology image classification while reducing the need for labelled data. Methods: We retrospectively analysed cytology images from 393 thyroid nodules and extracted 142,158 image patches. Three self-supervised learning methods were used to pre-train vision transformer models on unlabelled images, followed by classification training with different amounts of labelled data. Performance in classifying eight lesion types was assessed using five-fold cross-validation, precision–recall area under the curve, and macro-averaged F1 scores. Model performance was compared with ImageNet-pretrained models and four cytotechnologists. Results: Here we show that self-supervised learning maintains high classification performance when labelled training data are limited and consistently outperforms ImageNet pre-training. With 1% of the labelled data, the best self-supervised model achieves a precision–recall area under the curve of 0.899, compared with 0.707 for ImageNet pre-training. The diagnostic pipeline achieves an average class-specific accuracy of 0.94, exceeding that of the four participating cytotechnologists. Self-supervised image representations also form lesion-specific clusters without diagnostic labels. Conclusions: Self-supervised learning enables accurate classification of thyroid cytology images while substantially reducing the need for expert annotation. This approach may facilitate the development of more efficient artificial intelligence-assisted cytology systems, although validation using data from multiple institutions is needed.