by Christos Evangelou, MSc, PhD – Medical Writer and Editor
A new artificial intelligence (AI) tool has demonstrated high sensitivity for detecting colorectal cancer metastases in lymph node slides, including isolated tumor cells that experienced pathologists described as "easy to miss." The system comprises a two-stage classification and segmentation pipeline and was validated on more than 650 whole slide images (WSIs) collected at two independent institutions. A small pilot study showed that pathologists who used it expressed strong enthusiasm for incorporating it into routine practice.
The study was published in Cancer Medicine.
Study Rationale
Lymph node metastasis in patients with colorectal cancer influences treatment decisions and prognosis. Guidelines from The Royal College of Pathologists recommend examining at least 12 nodes per resection specimen, which can place a heavy burden on pathologists. In addition, early metastatic deposits can be small, and isolated tumor cells with weak glandular features closely resemble benign findings such as sinus histiocytosis.
"Lymph node assessment in colorectal cancer is clinically crucial but time-consuming and vulnerable to missed small metastases," said Alexey Fayzullin, PhD, researcher at Sechenov University's Institute for Regenerative Medicine and the study's first author, in an interview with Pathology News. "We wanted to test whether AI could support pathologists in this task."
The Two-Stage Pipeline
The research team designed a system that uses a two-stage pipeline. The first stage uses an image classification network, GoogleNet, to scan WSIs and identify tiles that may contain tumor tissue. This first stage serves as triage, as tiles with no suspicious cells are discarded, reducing the computational load required for closer inspection. The second stage runs the flagged tiles through DeepLabV3+, a segmentation model that produces pixel-level masks marking the borders of metastatic tissue.
Fayzullin explained that the two models working together let the system analyze images quickly while still generating the spatial information a pathologist needs.
"We developed an AI pipeline that first finds suspicious slide regions and then outlines metastatic tumor tissue," Fayzullin stated. "It was tested on WSIs from different institutions and scanners."
The training set consisted of 108 slides scanned on a Hamamatsu NanoZoomer S360. Slides used for training were annotated by three pathologists with at least four years of experience each, and annotations were validated by a senior pathologist with 14 years of practice. The model was then validated in 514 additional slides from the same scanner and 40 slides from a Leica Aperio AT2.
Key Findings
For the primary classification task of determining whether a WSI contains metastases, the model demonstrated 100% recall on the validation set of 514 slides. The system correctly identified all 102 slides that contained metastases and classified 385 of 412 negative slides correctly, yielding a specificity of 93.45%. The authors explained that the 27 false positives were largely attributable to tissue folds from microtome preparation.
The segmentation performance of the model was assessed using 70 slides with expert-annotated masks. The overall Dice coefficient was 0.818 ± 0.105 (0.841 ± 0.092 on the Hamamatsu dataset; 0.801 ± 0.112 on the Leica dataset).
For the segmentation task, the model accurately identified the boundaries of tumor cell clusters and excluded surrounding lymphoid stroma, mucin lakes, and necrotic tissue. The model did not generate false-positive masks in areas of sinus histiocytosis, a common source of diagnostic confusion in this setting.
What Pathologists Experienced
To assess real-world utility, the researchers created an online project within the Axon medical information system containing three cases. Two experienced pathologists reviewed 30 slides each, half in standard mode and half with AI-generated overlay masks. Time per slide was recorded on video.
There was no statistically significant difference in review time between the assisted and unassisted conditions (p = 0.23 for Expert 1; p = 0.17 for Expert 2). The overall median time per slide was around 21.5 seconds in the assisted condition.
"The model detected all metastatic cases in the validation set and accurately highlighted even very small tumor foci," Fayzullin said. "Unexpectedly, its main value was not speed, but diagnostic confidence."
In a slide containing a micrometastasis measuring 0.14 mm by 0.06 mm, both pathologists flagged it as something they would have been at risk of missing in routine practice. The lesion had weak glandular architecture, a feature that characterizes adenocarcinomas including colorectal cancer, and it closely resembled histiocytosis. The AI model included the micrometastasis in the segmentation mask.
Feedback overall was strongly positive; both experts expressed a desire to use the system in their day-to-day work.
Potential Implications
The immediate application the team envisions is to enhance the workflow for lymph node assessment for metastases.
"The main clinical application we envision is a background AI tool that automatically prioritizes lymph node cases with suspected metastases, helping pathologists focus their attention where it is most needed," Fayzullin said.
With 100% recall on metastatic slides, a validated version of the system could help a pathologist deprioritize negative slides. The researchers estimated this could reduce the number of slides a pathologist needs to examine per case by a factor of 10 for non-metastatic cases, although the authors emphasized that full diagnostic responsibility remains with the clinician.
Limitations and Future Work
The authors acknowledge that further validation is needed, as the pilot evaluation involved only two pathologists reviewing 30 slides each and the training dataset was from a single institution and a single scanner format. The potential time-saving benefit may become more apparent when the system is used by less experienced pathologists, who might take longer in the unassisted condition.
"Key unanswered questions concern how AI can move beyond detecting metastases toward interpreting their biological and clinical significance," Fayzullin said. "Future studies should combine multiple-instance learning and multimodal models to evaluate metastatic morphology and potentially predict metastatic involvement even when the tissue section does not pass directly through the metastatic focus."
This study was supported by the Ministry of Science and Higher Education of the Russian Federation.
References
- Fayzullin A, Savelov N, Balkivskiy A, et al. Region-Based Segmentation of Lymph Node Metastases in Whole-Slide Images of Colorectal Cancer: A Pilot Clinical Study. Cancer Med. 2026;15(3):e71449. doi:10.1002/cam4.71449
