PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Computational Cytology: A Survey

Hao Jiang, Yanning Zhou, Yi Lin, Ronald CK Chan, Jiang Liu, Hao Chen

arXiv:2202.05126v2eess.IVcs.CVcs.LG

TL;DR

Reviewing every cell for malignancy is time-consuming, and prior DL-based cytology surveys were not exhaustive or often focused on the cervix. This paper surveys more than 120 publications on DL methods and cytology applications across body sites. It summarizes progress and highlights challenges including burdensome annotation and limited generalization of simply designed CNNs.

  • Problem

    Reviewing every cell in cytology slides or gigapixel whole-slide images is time-consuming and tedious for cytologists.

    Source: §1 · p. 2

  • Method

    The survey reviews more than 120 publications on deep learning approaches and cytology specimens from various body parts.

    Source: §1 · p. 3

  • Results

    The survey covers supervised, weakly supervised, unsupervised, and transfer learning, and reviews classification, detection, segmentation, datasets, metrics, and applications across multiple cytology types.

    Source: §6 · p. 22

  • Takeaways & Limitations

    Deep learning-based methods show promise for efficient cancer screening, while the survey identifies current challenges and potential research directions.

    Source: §6 · p. 22

    Effective labeling processes for cytology images containing numerous cells, especially gigapixel whole-slide images, remain largely unexplored.

    Source: §5.1 · p. 20

Abstract

from arXiv · show

Computational cytology is a critical, rapid-developing, yet challenging topic in the field of medical image computing which analyzes the digitized cytology image by computer-aided technologies for cancer screening. Recently, an increasing number of deep learning (DL) algorithms have made significant progress in medical image analysis, leading to the boosting publications of cytological studies. To investigate the advanced methods and comprehensive applications, we survey more than 120 publications of DL-based cytology image analysis in this article. We first introduce various deep learning methods, including fully supervised, weakly supervised, unsupervised, and transfer learning. Then, we systematically summarize the public datasets, evaluation metrics, versatile cytology image analysis applications including classification, detection, segmentation, and other related tasks. Finally, we discuss current challenges and potential research directions of computational cytology.

1. Introduction

Cytology examines cellular morphology for cancer screening, but reviewing every cell in a slide or gigapixel whole-slide image is time-consuming. This survey reviews deep learning methods and applications across cytology specimens and identifies gaps in earlier reviews.

1. Introduction

  • Cytology studies cellular morphology and composition, usually for cancer screening, whereas histopathology focuses on cells within tissue architecture.Clinical specimen preparation includes collection and preservation, centrifugation, slide making, and staining.Source: §1 · pp. 1–2
  • Cytology specimens may be exfoliative, abrasive, or aspiration-based; unlike histology, they do not require removal of intact sizable tissue.The passage describes cytology sampling as less invasive, frequently painless, low-cost, and equipment-undemanding.Source: §1 · p. 1
  • Reviewing every cell under a microscope or in gigapixel whole-slide images is time-consuming and tedious for cytologists.Researchers have pursued automatic methods for accurate and efficient cancer screening as caseloads increase.Source: §1 · pp. 2–3
  • Deep learning learns feature representations end to end, reducing reliance on task-related features designed from expert knowledge in traditional machine learning.The survey notes applications across cervix, breast, bladder, and lung, with supervised learning the most commonly developed scheme.Source: §1 · p. 3
  • DL-based computational cytology research increased 10-fold from 2014 to 2021, with a booming trend in the two most recent years.The publications address classification, detection, segmentation, and other tasks.Source: §1 · p. 3, Figure 3 · p. 3
  • Earlier reviews were not exhaustive on advanced algorithms, public datasets, or trends, and most DL surveys focused on cervix rather than other cancers.This survey covers more than 120 publications since 2014 and cytology specimens from various body parts.Source: §1 · p. 3

2. Deep learning methodology

The survey organizes deep-learning methods by annotation availability into supervised, weakly supervised, unsupervised, and transfer-learning approaches.

2.1. Supervised learning

  • Supervised learning maps images to task-specific labels, with cell-level classification predicting patch categories and slide-level screening aggregating multilevel predictions into WSI results.Source: §2.1 · p. 4, Figure 4 · p. 4
  • Detection methods use either two-stage region proposals and refinement or direct prediction, while segmentation distinguishes semantic pixel labels from instance IDs.Mask R-CNN is an example of instance segmentation.Source: §2.1 · p. 5, Figure 5 · p. 5

2.2. Weakly supervised learning

  • Weak supervision covers incomplete, inaccurate, and inexact labels; multiple instance learning uses coarse annotations to learn fine-grained tasks through bags of instances.Source: §2.2 · p. 5, Figure 6 · p. 6

2.3. Unsupervised learning

  • Unsupervised methods learn from unlabeled data; cytology examples include stain conversion, classification augmentation, and GAN-based high-resolution image generation.VAE constrains latent variables to a normal distribution, while GANs use a generator and discriminator to produce realistic images.Source: §2.3 · p. 6, Figure 7 · p. 7

2.4. Transfer Learning

  • Transfer learning moves source-domain knowledge through fine-tuning or domain adaptation; fine-tuning pretrains on large task-similar data before training final layers on smaller target data.Domain adaptation aligns dataset distributions to transfer knowledge between domains.Source: §2.4 · pp. 6–7

3. Datasets and metrics

The survey catalogs public and representative private cytology datasets, noting a strong cervical focus and fewer datasets for other cancer types. It also summarizes metrics for classification, detection, and segmentation.

3. Datasets and metrics

  • Public cytology datasets are predominantly cervical, with smaller numbers covering cancers such as breast, oral, and stomach.The survey also summarizes representative private databases.Source: §3 · p. 7, Table 1 · p. 8
  • The survey covers public cytology datasets used for cervical classification and segmentation, nuclei detection, breast classification, and refocusing and super-resolution, with examples including Herlev, ISBI 2014/2015, and FNAC.Source: §3 · pp. 7, 10
  • Available datasets support varied tasks: oral images include category, box, and mask labels, while CDetector provides 48,587 boxes across 11 categories.The survey identifies CDetector as the largest public cytology object-detection dataset.Source: §3 · pp. 8, 10
  • Accuracy evaluates category prediction but ignores class imbalance; detection uses IoU and AP at thresholds such as AP50 and AP75, while segmentation uses pixel accuracy and Dice.Dice measures similarity between predicted and ground-truth masks.Source: §3 · p. 10

4. Deep learning in cytology application

This section surveys deep-learning models applied to computational cytology, covering preprocessing and clinical image-analysis tasks. Figure 8 illustrates several staining techniques used in cytology.

4. Deep learning in cytology application

  • The survey covers preprocessing and deep-learning applications in classification, detection, segmentation, and other related tasks.It presents representative studies across these areas.Source: §4 · p. 10
  • The surveyed work includes preprocessing techniques before its discussion of classification, detection, segmentation, and other tasks.Source: §4 · p. 10
  • Figure 8 presents six staining techniques: Pap, H&E, Giemsa, Feulgen, AgNOR, and Diff-quik.The figure labels examples of each technique.Source: Figure 8 · p. 10

4.1. Preprocessing

The survey discusses staining, stain normalization, and data augmentation as preprocessing topics in cytology image analysis. It describes staining's role in visualizing cellular structures and normalization's relevance to color variation.

Staining techniques

  • Pap, H&E, and Giemsa are common stains with different uses, while AgNOR, Feulgen, and other techniques appear in specific applications.Pap is especially common in cervical cytology; Giemsa is used for blood and bone-marrow evaluation, and AgNOR presents cell proliferation, differentiation, and malignant transformation.Source: §4.1 · p. 10

Stain Normalization

  • Stain variation from techniques and imaging conditions challenges robust model generalization, and normalization can accelerate network convergence.The survey names whitening and linear normalization as examples.Source: §4.1 · p. 10

Data augmentation

  • Augmentation expands limited image sets through geometric or color transformations and GAN-based synthesis.Yu et al. generated 16,000 images from 961 real images for cervical-cell classification; PGGAN-generated lung images were used to train classification CNNs with reported performance improvements.Source: §4.1 · pp. 10–11

4.2. Classification

The survey divides cytology classification into cell-level predictions and slide-level predictions, describing methods for both. It covers architectural adaptations, data limitations, and multi-stage whole-slide analysis.

4.2.1. Cell-level classification

  • Cell-level classification crops cell patches from gigapixel whole-slide images and feeds them to deep models after preprocessing.Straightforward CNN approaches extract feature maps and predict categories, but their performance and generalization may be limited to specific datasets and scenarios.Source: §4.2.1 · p. 11
  • Fine-grained cervical-cell classification uses morphological mask maps to strengthen distinctions between ambiguous sub-categories.The challenge is acute in multi-class settings where class boundaries can be ambiguous.Source: §4.2.1 · p. 11
  • GAN-based synthesis and sampling address limited or imbalanced classification datasets.Conditional GAN synthesis of 180 images, combined with original data, improved reported FNAC classification performance across ResNet-152, DenseNet-161, and Inception-V3.Source: §4.2.1 · pp. 11, 14
  • Attention, pyramid pooling, LSTM aggregation, and graph convolution are among advanced cell-classification designs surveyed.The studies use these components to focus on cell areas, aggregate regional features, or explore relationships among cervical-cell images.Source: §4.2.1 · p. 14

4.2.2. Slide-level classification

  • Slide-level classification predicts whole-image categories, and multi-stage systems combine local cell or patch analysis with whole-slide prediction.One reported transformer-based framework achieved a higher AUC score and was 2.51× faster than state-of-the-art methods in WSI classification.Source: §4.2.2 · p. 14
  • Slide-level approaches also use limited annotations, auxiliary tasks, and risk-stratification pipelines for whole-slide decisions.A reported cervical risk-stratification method was validated on 19,303 WSIs from multiple medical centers.Source: §4.2.2 · p. 15

4.3. Detection

Cytology detection locates and identifies small objects, with one-stage methods predicting boxes directly and two-stage methods proposing candidates before classification and localization. Studies illustrate a speed–accuracy trade-off between these approaches.

4.3. Detection

  • One-stage detectors directly regress object categories and locations, whereas two-stage detectors first propose candidates and then classify and localize them.The section characterizes one-stage methods as efficient; two-stage methods use strategies such as selective search or region proposal networks.Source: §4.3 · p. 15, §4.3.1 · p. 15, §4.3.2 · p. 15

4.3.1. One-stage methods

  • YOLO detected cervical cells and pleural-effusion nuclei at 0.060 second/image, compared with 1.627 second/image for Faster R-CNN.The survey also describes YOLO use on selected ROI tiles for bone-marrow object detection and classification, and a context-aware cervical detector that filters false positives.Source: §4.3.1 · p. 15
  • RetinaNet uses a feature pyramid network to learn multiscale features; its cytology application quantified pulmonary hemosiderophages.The detector generated multiscale features for box regression and classification.Source: §4.3.1 · p. 15

4.3.2. Two-stage methods

  • Two-stage cytology detectors include Faster R-CNN applications for cervical exfoliated cells and tumor cells in an ascites interpretation system.The section also discusses structured regression using dot annotations and weakly supervised or semi-supervised localization and detection strategies.Source: §4.3.2 · pp. 15–16
  • R-FCN achieved higher mAP while SSD tested faster; Faster R-CNN had better average precision than RetinaNet, while RetinaNet was significantly faster in limited-data cervical detection.These comparisons support a speed–accuracy trade-off among detection models.Source: §4.3.2 · p. 17

4.4. Segmentation

Cytology segmentation assigns image pixels to categories to delineate objects, supporting cell separation, cell-type distinction, and cellular-structure analysis. The surveyed approaches address overlapping cells either by refining coarse segmentations or by detecting and segmenting instances.

4.4. Segmentation

  • Segmentation is pixel-wise classification that delineates object contours and can separate cells from background, distinguish cell types, or delineate nuclei and cytoplasm.Overlapping areas between cells are identified as the principal segmentation challenge.Source: §4.4 · p. 17
  • Overlapping cells are addressed either by segmenting coarsely and refining boundaries, or by detecting objects and predicting instance masks in one architecture.The latter detect-then-segment approach avoids a separate overlap-refinement design.Source: §4.4 · p. 17

4.4.1. Segment-then-refine method

  • Segment-then-refine methods use a coarse segmentation, often U-Net, followed by strategies such as erosion, watershed, graph-based processing, or non-maximum suppression to separate overlaps.Refinement methods can improve segmentation in overlapping regions, while complex clinical data can challenge reproducibility and generalizability.Source: §4.4.1 · p. 17

4.4.2. Detect-then-segment method

  • Detect-then-segment methods detect objects across the image and segment each instance; Mask R-CNN can jointly output classification, detection, and segmentation.Cytology studies also report attention-enhanced instance representations and semi-supervised learning using labeled and unlabeled data.Source: §4.4.2 · pp. 18–19

4.5. Other tasks

Beyond classification, detection, and segmentation, cytology image analysis also includes super-resolution and stain conversion. These methods address image quality and staining heterogeneity.

4.5. Other tasks

  • Super-resolution models reconstruct high-resolution cytopathology images from low-resolution inputs, including multi-focal-plane images from a single 2D image.PathSRGAN maps real low- and high-resolution images; STSRNet reconstructs high-resolution multi-focal-plane images and refocuses them.Source: §4.5 · p. 19
  • CycleGAN-based style transfer converted Giemsa-stained images to Pap-stained images, with visual evaluation of synthetic lung-cell nuclei, cytoplasm, and layouts.Stain conversion is presented as a way to normalize staining and address data heterogeneity.Source: §4.5 · p. 19

5. Challenges and Promises

The survey identifies continuing challenges in computational cytology, including limited annotations, fine-grained analysis, generalization, interpretability, and clinical integration. It describes research directions spanning label-efficient learning, robust models, explanations, and human-AI collaboration.

5. Challenges and Promises

  • The survey frames computational cytology as advancing while still facing open challenges and research problems.Its discussion focuses on potential research directions for the field.Source: §5 · p. 19

5.1. Label-efficient learning with limited annotations

  • Public cytology datasets are limited in number, cancer types, and annotation types, while manual ROI and object labeling burdens cytologists.The survey discusses semi-supervised learning and multiple instance learning as ways to use limited annotations; efficient cytology labeling remains largely unexplored.Source: §5.1 · pp. 19–20

5.2. Fine-grained classification and morphological feature characterization

  • Fine-grained cell classification is harder than coarse classification, and overlapping cells complicate instance segmentation needed to calculate nuclear-cytoplasmic ratios.Multi-stage refinement designs can increase training difficulty and impair generalization; end-to-end instance segmentation remains an area for further study.Source: §5.2 · p. 20

5.3. E

  • Cytology models need effective feature representations; attention can focus on lesion-related regions, while transformer-based dependencies remain to be explored.The survey notes attention-based cytology studies in segmentation and instance representation, with reported improved performance.Source: §5.3 · p. 20

5.4. Generalizability and robustness

  • Specimen collection, staining, and imaging differences contribute to data heterogeneity, which can weaken robustness and generalization on unseen data.Normalization offers limited benefit in the described setting; domain adaptation and cross-domain cytology learning remain areas for exploration.Source: §5.4 · pp. 20–21

5.5. Transparency and interpretability

  • Clinical adoption is restricted by deep learning's black-box nature; proposed interpretability directions include slide-level screening rules and visualizing model decisions.The survey describes heatmaps, Grad-CAM, LIME, and multi-stage systems that expose intermediate analysis results.Source: §5.5 · p. 21

5.6. Digital medicine and human-AI collaboration

  • Clinical transformation is challenged by unavailable data, diverse clinical conditions, and differing diagnostic criteria across cancer types.The survey identifies privacy-preserving learning, robust architectures, and human-machine collaboration as areas relevant to clinical use; accurate and efficient diagnosis and system integration remain challenges.Source: §5.6 · pp. 21–22

6. Conclusion

The conclusion surveys deep-learning methods and applications in computational cytology, alongside datasets, evaluation metrics, challenges, and future research directions.

6. Conclusion

  • Deep learning-based cytology research spans supervised, weakly supervised, unsupervised, and transfer learning methods.Source: §6 · p. 22
  • The survey covers cytology image analysis applications in classification, detection, and segmentation across multiple organs.Examples include cervix, breast, lung, thyroid, oral, kidney, and stomach.Source: §6 · p. 22
  • The paper summarizes public datasets and evaluation metrics for developing new models.Source: §6 · p. 22
  • The review outlines current challenges and potential directions for future computational cytology research.Source: §6 · p. 22

CRediT authorship contribution statement

The statement assigns authors contributions spanning conceptualization, methods, analysis, investigation, writing, visualization, and project support.

CRediT authorship contribution statement

Related papers

Loading 2202.05126v2…