{"title":"Deep learning approaches for pathological image classification","authors":"Masayuki Tsuneki","doi":"10.1016/j.job.2025.100696","DOIUrl":null,"url":null,"abstract":"<div><h3>Background</h3><div>Computer-aided diagnosis using deep learning has emerged as a transformative tool in pathology that enables the automated and consistent interpretation of whole slide images. Among the key tasks in pathological image analysis, classification-based deep learning models have shown promise in distinguishing cancer subtypes and predicting genetic or molecular features. However, challenges remain due to the limited availability of high-quality labeled datasets, particularly for rare cancers, which hinders the widespread applicability of conventional data-driven approaches.</div></div><div><h3>Highlight</h3><div>The methodology for developing classification-based deep learning models involves data preparation, annotation strategies, training techniques, and performance evaluation. Core supervised learning approaches are used alongside convolutional and recurrent neural networks to enhance classification performance. In scenarios in which training datasets are scarce, transfer learning serves as an effective strategy to improve the model training efficiency. In addition, this review introduces emerging techniques such as the use of simulators to generate synthetic data and formula-driven approaches that simulate realistic scenarios, with the aim of overcoming the limitations of conventional training datasets. Visualization tools such as probability heatmaps are critical for model interpretability and validation.</div></div><div><h3>Conclusion</h3><div>Classification-based deep learning models are expected to play an increasing role in precision medicine by expanding capabilities beyond diagnosis to include prognostic prediction and treatment decision making. With innovations in augmented intelligence, deep learning models can comprehensively support clinicians across the diagnostic and therapeutic spectrum, particularly in an era of increasing demand and limited human resources in pathology.</div></div>","PeriodicalId":45851,"journal":{"name":"Journal of Oral Biosciences","volume":"67 4","pages":"Article 100696"},"PeriodicalIF":2.3000,"publicationDate":"2025-09-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Oral Biosciences","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1349007925000854","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"DENTISTRY, ORAL SURGERY & MEDICINE","Score":null,"Total":0}
引用次数: 0
Abstract
Background
Computer-aided diagnosis using deep learning has emerged as a transformative tool in pathology that enables the automated and consistent interpretation of whole slide images. Among the key tasks in pathological image analysis, classification-based deep learning models have shown promise in distinguishing cancer subtypes and predicting genetic or molecular features. However, challenges remain due to the limited availability of high-quality labeled datasets, particularly for rare cancers, which hinders the widespread applicability of conventional data-driven approaches.
Highlight
The methodology for developing classification-based deep learning models involves data preparation, annotation strategies, training techniques, and performance evaluation. Core supervised learning approaches are used alongside convolutional and recurrent neural networks to enhance classification performance. In scenarios in which training datasets are scarce, transfer learning serves as an effective strategy to improve the model training efficiency. In addition, this review introduces emerging techniques such as the use of simulators to generate synthetic data and formula-driven approaches that simulate realistic scenarios, with the aim of overcoming the limitations of conventional training datasets. Visualization tools such as probability heatmaps are critical for model interpretability and validation.
Conclusion
Classification-based deep learning models are expected to play an increasing role in precision medicine by expanding capabilities beyond diagnosis to include prognostic prediction and treatment decision making. With innovations in augmented intelligence, deep learning models can comprehensively support clinicians across the diagnostic and therapeutic spectrum, particularly in an era of increasing demand and limited human resources in pathology.