Transfer learning for medical image classification: a literature review.
Kim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME, Ganslandt T. · 2022
Read Transfer learning for medical image classification: a literature review. by Kim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME, Ganslandt T. (2022) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc9007400.txt.
Introduction
Kim and co-authors review transfer learning for sorting medical images, and guide readers on model choice. Transfer learning reuses what a network learned on other tasks. It eases the shortage of labeled medical images. It also saves time and computing power. Yet most studies set it up by habit, not by test. The team searched two databases for peer-reviewed work. They screened 425 papers in all. Of those, 121 fit the review's scope.
Most studies tried several models head to head. Deep models beat shallow ones in the tally. Inception was the single most used deep model. Teams also tested transfer styles against each other. Using the network as a fixed feature reader was the most favored single style. Training from scratch with a head start came second. The authors urge deep models as feature readers. That route saves cost without losing accuracy.
Catalog record
| CREATOR | Kim HE, Cosa-Linan A, Santhanam N, Jannesari M, Maros ME, Ganslandt T. |
|---|---|
| YEAR | 2022 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Transfer learning for medical image classification: a literature review By Hee E Kim, Alejandro Cosa-Linan, Nandhini Santhanam, Mahboubeh Jannesari, Mate E Maros, Thomas Ganslandt Published in: BMC Medical Imaging DOI: 10.1186/s12880-022-00793-7 ABSTRACT Abstract Background Transfer learning (TL) with convolutional neural networks aims to improve performances on a new task by leveraging the knowledge of similar tasks learned in advance. It has made a major contribution to medical image analysis as it overcomes the data scarcity problem as well as it saves time and hardware resources. However, transfer learning has been arbitrarily configured in the majority of studies. This review paper attempts to provide guidance for selecting a model and TL approaches for the medical image classification task. Methods 425 peer-reviewed articles were retrieved from two databases, PubMed and Web of Science, published in English, up until December 31, 2020. Articles were assessed by two independent reviewers, with the aid of a third reviewer in the case of discrepancies. We followed the PRISMA guidelines for the paper selection and 121 studies were regarded as eligible for the scope of this review. We investigated articles focused on selecting backbone models and TL approaches including feature extractor, feature extractor hybrid, fine-tuning and fine-tuning from scratch. Results The majority of studies (n = 57) empirically evaluated multiple models followed by deep models (n = 33) and shallow (n = 24) models. Inception, one of the deep models, was the most employed in literature (n = 26). With respect to the TL, the majority of studies (n = 46) empirically benchmarked multiple approaches to identify the optimal configuration. The rest of the studies applied only a single approach for which feature extractor (n = 38) and fine-tuning from scratch (n = 27) were the two most favored approaches. Only a few studies applied feature extractor hybrid (n = 7) and fine-tuning (n = 3) with pretrained models. Conclusion The investigated studies demonstrated the efficacy of transfer learning despite the data scarcity. We encourage data scientists and practitioners to use deep models (e.g. ResNet or Inception) as feature extractors, which can save computational costs and time without degrading the predictive power. Supplementary Information The online version contains supplementary material available at 10.1186/s12880-022-00793-7. Keywords: Deep learning, Transfer learning, Fine-tuning, Convolutional neural network, Medical image analysis INTRODUCTION Medical image analysis is a robust subject of research, with millions of studies having been published in the last decades. Some recent examples include computer-aided tissue detection in whole slide images (WSI) and the diagnosis of COVID-19 pneumonia from chest images. Traditionally, sophisticated image feature extraction or discriminant handcrafted features (e.g. histograms of oriented gradients (HOG) features [1] or local binary pattern (LBP) features [2]) have dominated the field of image analysis, but the recent emergence of deep learning (DL) algorithms has inaugurated a shift towards non-handcrafted engineering, permitting automated image analysis. In particular, convolutional neural networks (CNN) have become the workhorse DL algorithm for image analysis. In recent data challenges for medical image analysis, all of the top-ranked teams utilized CNN. For instance, the top-ten ranked solutions, excepting one team, had utilized CNN in the CAMELYON17 challenge for automated detection and classification of breast cancer metastases in whole slide images [3]. It has also been demonstrated that the features extracted from DL surpassed that of the handcrafted methods by Shi et al. [4]. However, DL algorithms including CNN require—under preferable circumstances—a large amount of data for training; hence follows the data scarcity problem. Particularly, the limited size of medical cohorts and the cost of expert-annotated data sets are some well-known challenges. Many research endeavors have tried to overcome this problem with transfer learning (TL) or domain adaptation [5] techniques. These aim to achieve high performance on target tasks by leveraging knowledge learned from source tasks. A pioneering review paper of TL was contributed by Pan and Yang [6] in 2010, and they classified TL techniques from a labeling aspect, while Weiss et al. [7] summarized TL studies based on homogeneous and heterogeneous approaches. Most recently in 2020, Zhuang et al. [8] reviewed more than forty representative TL approaches from the perspectives of data and models. Unsupervised TL is an emerging subject and has recently received increasing attention from researchers. Wilson and Cook [9] surveyed a large number of articles of unsupervised deep domain adaptation. Most recently, generative adversarial networks (GANs)-based frameworks [10–12] gained momentum, a particularly promising approach is DANN [13]. Furthermore, multiple kernel active learning [14] and collaborative unsupervised methods [15] have also been utilized for unsupervised TL. Some studies conducted a comprehensive review focused primarily on DL in the medical domain. Litjens et al. [16] reviewed DL for medical image analysis by summarizing over 300 articles, while Chowdhury et al. [17] reviewed the state-of-the-art research on self-supervised learning in medicine. On the other hand, others surveyed articles focusing on TL with a specific case study such as microorganism counting [18], cervical cytopathology [19], neuroimaging biomarkers of Alzheimer's disease [20] and magnetic resonance brain imaging in general [21]. In this paper, we aimed to conduct a survey on TL with pretrained CNN models for medical image analysis across use cases, data subjects and data modalities. Our major contributions are as follows: - (i)An overview of contributions to the various case studies is presented; - (ii)Actionable recommendations on how to leverage TL for medical image classification are provided; - (iii)Publicly available medical datasets are compiled with URL as a supplementary material. The rest of this paper is organized as follows. Section 2 covers the background knowledge and the most common notations used in the following sections. In Sect. 3, we describe the protocol for the literature selection. In Sect. 4, the results obtained are analyzed and compared. Critical discussions are presented in Sect. 5. Finally, we end with a conclusion and the lessons learned in Sect. 6. Figure 1 is the main diagram which presents the whole manuscript. [Figure: Visual abstract summarizing the scope of our study] BACKGROUND Transfer learning Transfer learning (TL) stems from cognitive research, which uses the idea, that knowledge is transferred across related tasks to improve performances on a new task. It is well-known that humans are able to solve similar tasks by leveraging previous knowledge. The formal definition of TL is defined by Pan and Yang with notions of domains and tasks. “A domain consists of a feature space X and marginal probability distributionP(X), whereX={x1,...,xn}∈X. Given a specific domain denoted byD=X,P(X), a task is denoted by T= Y,f(·) where Y is a label space and f(·) is an objective predictive function. A task is learned from the pair {xi,yi} where xi∈X andyi∈Y. Given a source domain DS and learning taskTS, a target domain DT and learning taskTT, transfer learning aims to improve the learning of the target predictive function fT(·) in DT by using the knowledge in DS andTS” [6]. Analogously, one can learn how to drive a motorbike TT (transferred task) based on one’s cycling skill Ts (source task) where driving two-wheel vehicles is regarded as the same domain DS = DT. This does not mean that one will not learn how to drive a motorbike without riding a bike, but it takes less effort to practice driving the motorbike by adapting one’s cycling