Breast Cancer Screening Using Lower-Resolution Imaging Could Help Reduce Pathologist Workload

by Christos Evangelou, MSc, PhD – Medical Writer and Editor

In a recent study, researchers at the Wellcome/EPSRC Centre for Interventional and Surgical Sciences and the Department of Computer Science at University College London investigated the minimum spatial and digital sampling resolution required for accurate binary classification of breast histology images into benign or malignant categories.

Using deep learning algorithms trained on public datasets of H&E-stained images, they found that macroscopic features visible even in low-resolution images were sufficient for this preliminary benign versus malignant discrimination.1

Specifically, researchers found that although resolutions below those of most commercial whole-slide imaging systems resulted in considerable image degradation, the quality of the resulting images was sufficiently high for deep learning models to accurately classify them as malignant or benign.1

These findings suggest that a low-cost, low-resolution imaging system paired with an accurate classifier could potentially serve as a pre-screening tool to reduce the pathologists’ workload.

“I believe the results of the study show that there is a real potential for a low-resolution ‘triage’ automated imaging system that could quickly remove benign histopathology samples from the queue without any human assessment,” said Lydia Neary-Zajiczek, PhD, who is the first author of the study.

She added that such an imaging system would greatly help with backlogs and the current staffing crisis in pathology and would allow pathologists more time to devote to research interests or more difficult cases.

The report was published in Medical Image Analysis.

 

Study Rationale: Addressing the Growing Workload of Pathologists

Cytology and histopathology services are indispensable components of cancer care pathways. However, steadily rising case volumes coupled with stagnant staffing levels have created an unprecedented crisis in cellular pathology departments worldwide.

In the UK, vacancy rates for consultant posts currently hover at 12.5%. The situation is equally concerning in other high-income nations as well as in low- and middle-income countries. Planned expansions of existing screening programs, such as for breast and colon cancer, will only further overburden the strained workforce.

Healthcare bodies have widely endorsed digitization of microscopy workflows as a crucial approach to increase efficiency. However, concerns about image quality and high equipment costs have slowed the adoption of digital pathology solutions.

“I first considered the fundamental question of this study after attending the 2019 European Congress on Digital Pathology and listening to the keynote speech by Jo Martin, then-president of the Royal College of Pathologists. Her talk addressed the clear enthusiasm in the pathology community to ‘go digital’ but identified a serious gap between this enthusiasm and the reality of the progress made in digitization up to that point,” Dr. Neary-Zajiczek noted.

She added that an expert pathologist presented cases in which he was able to classify a slide at a high level (i.e., healthy vs. potentially cancerous) very quickly and at very low magnification. Nonetheless, his interaction with a typical digital slide on his NHS-issued workstation was much slower, as the file was very large and the Internet connection in the building was poor.

“My background is primarily in optics, and I wondered if this rapid diagnosis was possible with a lower-resolution image. I found that there was a gap in the literature for a quantitative investigation of this problem,” Dr. Neary-Zajiczek said.

 

Investigating Imaging Resolution Requirements

The researchers systematically analyzed how reducing the spatial resolution of digitized histology slides impacts diagnostic accuracy. They computationally degraded several public datasets of H&E-stained breast tissue images with known benign and malignant labels. Imaging parameters, such as the numerical aperture and pixel size, were estimated for each dataset.1

“To answer whether a rapid diagnosis with a low-resolution image is possible, we combined an objective assessment of slide image quality based on the spatial resolution of the imaging system with a trained machine learning algorithm with comparable accuracy to an expert pathologist to estimate the minimum image quality needed for accurate classification of breast histology images,” Dr. Neary-Zajiczek noted.

A deep learning ensemble model consisting of VGG19, MobileNetV2, and DenseNet201 was trained on the original high-resolution images to establish baseline classification accuracy.1 The ensemble deep learning model achieved an accuracy of over 95% in classifying the original high-resolution images.

The images were then computationally degraded to a lower resolution by modulating the optical transfer function. The pre-trained models and baseline classifiers were tested on these degraded versions of the validation set with progressively lower numerical aperture and digital sampling frequency to determine the threshold at which the accuracy dropped significantly.

“This degradation model is, to my knowledge, the first attempt to link blurriness, a subjective but very important image quality metric, with numerical aperture, a real-world parameter that is readily available in all imaging systems,” Dr. Neary-Zajiczek explained.

All previous studies used subjective assessments of image quality and linked these to diagnostic confidence, whereas this study provided quantitative measures of both image quality and accuracy.

 

Surprisingly Robust Performance

Across all image sets, the classifiers remained remarkably robust as the images were reduced to significantly lower resolutions and macroscopic morphological features were sufficient for binary classification.1 For example, the model trained on the PatchCamelyon dataset achieved 83% accuracy even when the numerical aperture was decreased from 0.13 to 0.09.

These findings suggest quantitatively confirm that high spatial frequency information is not strictly necessary for the classification of breast histology images as malignant or benign.

The models re-trained on low-resolution images performed even better. For example, the model re-trained on the PatchCamelyon dataset achieved 86% accuracy even when the numerical aperture was decreased from 0.13 to 0.04.

For the BreaKHis images acquired at 4X magnification, the re-trained model maintained 97% accuracy when the numerical aperture was reduced from 0.16 to 0.04. This corresponds to switching from a 10X to a 1.25X microscope objective.

 

Practical Implications

The findings of this study suggest that a rapid digital imaging device using low-power objectives could generate sufficient imaging data to serve as an initial screening step.

By automatically identifying normal slides that require no further examination, such a system could reduce caseloads for pathologists. This would allow them to allocate more time and resources to more complex tasks, such as grading malignancies.

In a scenario in which false negatives are unacceptable, optimizing for high specificity rather than outright accuracy could be advantageous. This approach could help develop an accurate pre-screening tool to alleviate workload pressures rather than replace pathologists.

Reduced scanning times and file sizes are additional benefits of lower-resolution digitization. Imaging at 4X versus 20X would require 25 times fewer images to cover the same slide area. Smaller data files would also reduce digital storage costs.

 

Study Limitations and Future Directions

Because this preliminary study relied solely on public datasets of breast tissue images, evaluating the approach in real-world clinical settings with fresh tissue samples is an important next step. Extending the analysis to other cancer types beyond breast could broaden its utility.

“With the inevitable introduction of more benchmark datasets for different tissue types, we can investigate if these results are applicable to other tissue types associated with high-volume screening programs, such as colon and prostate tissues,” Dr. Neary-Zajiczek said.

In addition, this study focused on image resolution as this was identified as the most important image quality metric; however, color balance and other image parameters are also very important and could be similarly quantified and compared using this experimental methodology.

 

References

  1. Neary-Zajiczek L, Beresna L, Razavi B, Pawar V, Shaw M, Stoyanov D. Minimum resolution requirements of digital pathology images for accurate classification. Med Image Anal. 2023;89:102891. doi:10.1016/j.media.2023.102891

Share This Post

Leave a Reply