Back to all blog posts
The new diagnostic categories of pulmonary adenocarcinoma
Pulmonary adenocarcinoma, also known as lung adenocarcinoma, is a non-small cell lung cancer (NSCLC). Before the release of the 2015 WHO classification of lung tumors, the classification of lung adenocarcinoma was much simpler. However, the 2015 classification introduced some new and important diagnostic categories – adenocarcinoma in situ and minimally invasive adenocarcinoma. While most lung adenocarcinomas are aggressive malignancies, these two new categories have an excellent prognosis. They are part of the “lepidic-predominant” category, so the size of the invasion does not necessarily correspond to the size of the whole tumor.
The introduction of these diagnostic categories made it important to measure the invasive component of the tumor. These invasive measurements determine the classification and T stage of the tumor. That was never necessary in the past when the gross tumor size (i.e., the size measured with a ruler) was the determinant of stage for all types of adenocarcinoma.
Pattern-based grading has also been introduced as an important concept in pulmonary adenocarcinoma over the past 20 years. Starting with the 2015 WHO, it was recommended that the predominant growth pattern be reported for all adenocarcinomas, along with listing all other patterns present in 5% increments. Since at least six growth patterns can be recognized, it is hard for pathologists to always agree on which predominant pattern they assign. But this is important for prognosis.
Challenges in pathological reporting
Pathological reporting related to lung adenocarcinoma may be challenging, particularly in determining invasion size in lepidic-predominant tumors. Significant interobserver variability exists among pathologists when determining invasive size in lung adenocarcinoma (PMID 27067781, 31107973) and the predominant invasive pattern (PMID 36343368, 22408209). AI algorithms could potentially help with these problems.
Tumor invasive size is an important parameter because it is the determining factor for the T stage in the TNM staging classification (AJCC Cancer Staging Manual 8th edition) of pulmonary adenocarcinoma. For many tumors, that is simply the gross measurement of the whole tumor mass, but for an important subset, the “lepidic-predominant” tumors, only a subset of the tumor is invasive. So, knowing how to measure that invasive component and being consistent in doing so are very important.
Dr. Jennifer M. Boland started the project to help pathologists with time-consuming, tedious tasks that suffer from interobserver variability when assessing pulmonary adenocarcinoma. “These factors are important to provide accurate staging and prognostic information for our patients but are difficult for us to do in an efficient and reproducible fashion,” she comments.
Designing and creating the AI model
The project team selected one representative hematoxylin and eosin (H&E) slide from 100 resected pulmonary adenocarcinomas. They were arbitrarily divided into training (n=35) and validation (n=65) sets. The slides were scanned and uploaded to Aiforia® Create for AI model creation.
Six expert pulmonary pathologists completed the annotations that were used to create a nested AI model. A pulmonary pathologist performed manual measurements using a digital ruler to calculate the largest tumor extent, largest invasive size, and invasion percentage. These values were then compared to those generated by the AI model.
For more example images, see Dr. Boland’s presentation from USCAP 2024.
How did Aiforia support the research?
Aiforia provided comprehensive support for the project team in developing the AI model. This included:
- Intensive model development meetings guiding Dr. Boland through the process of refining AI model parameters such as field of view (FOV) and iteration counts to enhance model performance.
- Suggestions for improving annotation quality to reduce white space and ensure more accurate model training outcomes.
- Technical advice on training both layers of the model (tumor vs. benign lung and tumor subregions of invasive vs. non-invasive) to optimize the distinction capabilities of the AI model.
- Assistance in troubleshooting and iterating the model, including advice on iteration numbers and the complexity setting for the model to ensure nuanced detection of tumor types.
The collaboration enhanced the model's accuracy in identifying invasive tumor regions, a critical aspect of the project. The efforts to refine annotation strategies and model parameters underscored Aiforia's commitment to supporting innovative research in digital pathology and expertise in successfully applying AI technologies to complex medical challenges.
“It has been a very fun collaboration. I have gotten to learn brand new things about AI, and I have been able to provide my expertise in pulmonary pathology to hopefully make any algorithms we design as good as possible. The Aiforia team has been great about providing technical suggestions, advice, and support.” – Dr. Boland, the Mayo Clinic
Read also: How did the Mayo Clinic choose a vendor for AI in pathology?
Results: the pathologist leads – AI assists
The AI model was successfully designed using expert pathologist annotations to aid in the assessment of pulmonary adenocarcinoma invasion. However, a manual pathologist’s review and potential revision of the model assessments are necessary to ensure accuracy.
The final model evaluated whole slide images of the training and validation sets. Invasion assessments by the AI model were compared to those manually measured or estimated by a pulmonary pathologist.
Most invasive percentages generated by the AI model were within 10% of pathologist assessment, with tumor and invasive measurements typically within 1-2 mm.
Dr. Boland and her team are happy with the model's ability to help identify areas of invasion in lung adenocarcinoma. "Our next step is to try and train a model to identify all the different pattern subtypes of invasive adenocarcinoma, which can be used to help assign tumor grade. Down the road, the model could potentially be trained to help in other areas of difficulty, including determination of pleural invasion and spread through alveolar spaces. I am excited to see what we can do," Dr. Boland concludes.
Read more lung cancer case studies:
- Pulmonary tumors case study: Ki-67 proliferation index calculation by AI
- MIT case study: advancing lung cancer research with AI
- PD-L1 case study: interview with Liesbeth Hondelink
References
Dr. Jennifer M. Boland from the Mayo Clinic and Lucas Stetzik, Aiforia’s Senior Scientist, were interviewed for this article.