A cancer misdiagnosis can delay appropriate treatment and potentially lead to poorer outcomes. But could artificial intelligence (AI) help support diagnostic testing and identify cases that warrant further investigation?
Board-certified oncologist and hematologist Vitor Pastorini encountered such a case when the clinical presentation of a patient with a presumed breast cancer diagnosis did not fully align with the initial findings. He ordered comprehensive molecular profiling that incorporated an AI-powered tissue-of-origin model to investigate further. Here, he discusses the case and the potential role of AI-supported tools in improving diagnostic accuracy.
Can you walk us through this case?
The patient was a 42-year-old woman whose screening mammogram identified a mass in the upper inner left breast. Ultrasound also showed multiple enlarged, bulky lymph nodes in the left axilla, and computed tomography (CT) revealed extensive bulky adenopathy throughout the mesentery and retroperitoneum.
Biopsy of the breast mass showed sclerosing adenosis, a benign finding. However, a concurrent axillary biopsy revealed a poorly differentiated malignancy that was estrogen receptor, progesterone receptor, and human epidermal growth factor receptor 2 negative and was initially diagnosed as grade 3 invasive ductal carcinoma. The tissue was subsequently sent for molecular profiling. Diagnostic immunohistochemistry (IHC) was not performed during the initial workup.
The combination of a benign breast biopsy and widespread bulky lymphadenopathy was clinically unusual. Imaging also identified an 8-cm multicystic mass in the posterior pelvis, raising the possibility of two separate primary malignancies.
What prompted you to use AI-assisted molecular profiling, and how did it change your diagnostic thinking?
AI-assisted molecular disease classifiers use exome and transcriptome sequencing data generated during tissue molecular profiling to help predict a histopathologic diagnosis. In this case, once the molecular data became available, the algorithm identified a discrepancy between the predicted and submitted diagnoses.
This finding prompted a diagnostic IHC workup to investigate the possibility of a hematologic malignancy. The diagnosis was quickly confirmed and communicated to me.
Why didn't the correct diagnosis emerge during the initial diagnostic workup?
The correct diagnosis was not initially identified because a hematologic malignancy was not considered during the original workup. In cases such as this, molecular profiling can provide diagnostic context beyond morphology, including the detection of diagnostic fusions, mutations, RNA expression patterns, viral infections such as HPV, EBV, MCV, and HHV8, and genomic signatures associated with exposures such as UV radiation or smoking.
Without this broader molecular data set – and without diagnostic IHC during the initial workup – the pathway to the correct diagnosis was more limited.
What does this case tell us about the role of AI-assisted molecular profiling in pathology?
This case illustrates how AI can integrate large volumes of molecular data and generate reproducible probabilities that pathologists can incorporate into their clinical judgment. Importantly, these tools are not generative AI. They are locked algorithms, typically based on neural networks, that undergo analytical and clinical validation prior to deployment.
AI-assisted molecular profiling may provide clinical utility in several settings, particularly in cases with ambiguous morphology, unusual clinical presentations, or overlapping differential diagnoses. One example is distinguishing primary lung squamous cell carcinoma from squamous cell carcinoma metastatic to the lung – a distinction with important treatment implications. Recent published research has demonstrated the potential for AI-assisted approaches to contribute to this differential diagnosis.
How do you see AI reshaping diagnostic workflows over the next few years?
Pathology is facing a widening gap between workforce capacity and case volume. As demands on pathologists increase, new tools may help improve efficiency. AI has the potential to reduce time spent on lower-complexity, repetitive tasks, such as counting mitotic figures on H&E slides, allowing pathologists to focus on complex, judgment-driven work that requires their expertise.
Cases such as this also underscore the importance of maintaining a broad differential diagnosis and considering all available data – including molecular findings and AI-derived signals – before reaching a final diagnosis.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
