Conventional pathological examination is considered the gold standard for nodal assessment; however, micrometastatic disease can often be missed during routine slide review, particularly in busy clinical settings. Missed lymph node metastases can lead to understaging and inappropriate treatment decisions, underscoring the need for more sensitive and efficient diagnostic tools.

As part of a multicenter diagnostic study, Shaoxu Wu, MD, of the Department of Urology at Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, China, and colleagues, developed PanCAM, a pan-cancer AI-driven diagnostic model designed to detect lymph node metastases across multiple cancer types. Findings from the retrospective and prospective validation of PanCAM were published in The Lancet Digital Health.

Study Details and Model Architecture 

The study included 9,256 patients who had undergone tumor resection and lymph node dissection from 17 hospitals in China between January 2013 and November 2024, which included 1,303 patients in the training set, 558 in the internal validation set, 6,006, in the external validation sets, and 1,389 in the prospective validation sets. The total data set encompassed 33 cancer types (9 common malignancies and 24 rare cancers), generating more than 69,000 whole-slide pathology images and nearly 154,000 lymph nodes for analysis. The investigators also incorporated 399 images from the publicly available CAMELYON16 data set from the Netherlands as the international validation cohort. The primary endpoint was diagnostic sensitivity for lymph node metastasis detection.

The AI component of the study centered on supervised deep learning using a segmentation framework that combined DeepLabv3+ with the RegNet-Y40 encoder. The investigators used pixel-level annotations created and reviewed by experienced pathologists to train the model to recognize metastatic tumor regions. To improve efficiency, the researchers implemented a model-assisted annotation strategy in which preliminary tumor regions identified by the algorithm were reviewed and corrected by pathologists before being added back into the training data set for iterative refinement. The model was trained incrementally, beginning with prostate cancer images and progressively incorporating data from eight additional common cancers.

For analysis, whole-slide images were processed through a sliding window to generate patches for the model input, which were then visualized through segmentation heatmaps and pixel-level prediction probabilities to classify the images as either positive or negative. Diagnosis determinations were made based on pixel threshold and the number of positive pixels threshold, with optimal values established by maximizing the F2 classification score, which prioritizes recall over precision to determine performance for more safety-critical systems. 

Comparisons were also made against other cancer-specific diagnostic models, including a model previously validated for bladder cancer and another for prostate cancer. 

Particular emphasis was placed on real-world validation. In the prospective phase, pathologists and PanCAM independently reviewed digitized pathology slides while remaining blinded to each other’s assessments. Discordant cases were adjudicated by senior pathologists, with immunohistochemistry used when necessary. This prospective design was intended to mimic clinical deployment and evaluate whether the system could effectively complement existing pathology workflows.

Key Results

PanCAM demonstrated consistently strong performance across institutions and cancer types. In retrospective validation, sensitivity ranged from 0.97 (95% confidence interval [CI] = 0.92–0.99) to 1.00 (95% CI = 0.98–1.00) across participating hospitals, while prospective validation sensitivity ranged from 0.93 (95% CI = 0.78–0.99) to 1.00 (95% CI = 0.98–1.00). The model achieved a sensitivity of 0.98 for rare cancers in both retrospective and prospective analyses, despite being trained only on common cancers. It also performed well in detecting micrometastases, an area in which conventional pathology frequently encounters difficulty.

Across validation cohorts, PanCAM identified 141 patients with lymph node metastases that had been missed by pathologists, the majority of which were micrometastatic lesions. In the prospective study alone, the system identified 21 additional positive cases overlooked during routine review. Although the model generated some false-positive findings, the authors reported that these required only a brief additional review time and were considered an acceptable tradeoff for reducing missed diagnoses.

Compared with cancer-specific models, PanCAM showed similar sensitivity and area under the curve values to the bladder cancer and prostate cancer models, but PanCAM showed improved positive predictive value and specificity compared with the bladder cancer–specific model.

The authors concluded: “PanCAM provided a generalizable solution for detecting lymph node metastasis across cancer types. With high sensitivity and robust performance, the model could assist pathologists in diagnosing lymph node metastasis, improving diagnostic accuracy, supporting treatment decision-making, and ultimately enhancing patient outcomes.” They also suggested that its robust performance and generalizability make the model potentially useful for adoption in settings with poor or fewer medical resources. 

DISCLOSURES: The study was funded by the National Natural Science Foundation of China and others. The authors all declared no conflicts of interest. For select code availability, visit thelancet.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.