Five Questions for an AI Pathologist

In the earlier part of this feature, I spoke directly with my pathology AI assistant, “Pathrick,” to discuss the realistic applications of machine learning for a laboratory professional. In this section, I’ll be putting Pathrick through his paces, determining just how useful AI – specifically, ChatGPT – currently is when faced with pathological image analysis. 

Due to current ChatGPT functionality, the link for threads with images cannot be shared at this time.

Let the quiz begin!

Question 1: Can you tell me what type of tissue this is and explain why?

Pathrick’s Answer: The image displays a histological section of what appears to be lung tissue, more specifically normal pulmonary parenchyma. Here are the features supporting this conclusion:

1. Alveolar Architecture: The most distinctive feature of lung tissue on histology is its spongy appearance due to the presence of numerous alveoli. Alveoli are the tiny air sacs in the lungs where gas exchange occurs
2. Thin-walled Alveole: The walls of the alveoli are very thin, facilitating efficient gas exchange between the air in the alveoli and the blood in the surrounding capitation
3. Presence of Alveolar Septa: These are the thin partitions between alveoli. Within these septa, you can often find capillaries and occasionally some alveolar macrophages
4. Alveolar Macrophages: Scattered throughout the alveoli are cells that appear slightly darker than the surrounding tissue These are likely alveolar macrophages, which are responsible for phagocytosing inhaled particles and pathogens
5. Lack of Glands or Mucus: The absence of features like mucus and glands helps differentiate this from bronchial or bronchiolar tissue
6. Capillaries: Within the thin alveolar septa, you can observe small blood vessels which are the pulmonary capillaries

In conclusion, the tissue in the image is lung tissue tissue, specifically representing the alveolar structures of the pulmonary parenchyma.

---

Question 2: I found something that looks like this in the sections – is this cancer?

Pathrick’s Answer: The image depicts glandular epithelial tissue resembling intestinal mucosa, characterized by crypts of Lieberkühn, columnar epithelial cells, and goblet cells. However, observed features like nuclear atypia, increased mitotic figures, and disrupted glandular architecture suggest possible neoplasia or dysplasia. While these characteristics may be indicative of a pre-malignant or malignant process, a definitive diagnosis requires comprehensive evaluation, considering clinical history and additional stains. It’s crucial to consult with a senior pathologist for an accurate assessment.

Matthew: Could this be a primary tumor of the lung? If so, would I classify it as a mucinous or non-mucinous tumor?

Pathrick’s Answer: Based on the image provided and the histological features observed, it’s plausible to consider that this could represent a primary tumor of the lung, specifically adenocarcinoma, which is one of the most common histological types of non-small cell lung cancer. 

To differentiate between mucinous (formerly known as mucin-producing) and non-mucinous subtypes of adenocarcinoma:

• Mucinous Adenocarcinoma: This subtype has cells that produce an abundance of mucin which often appears as extracellular, pale blue-to-clear material on standard H&E staining. The tumor cells tend to float within the mucin, giving a “floating in lakes” appearance. The presence of goblet cells and copious amounts of mucin are hallmarks of this subtype
• Non-mucinous Adenocarcinoma: This subtype does not produce significant extracellular mucin, and the tumor architecture can be varied, including acinar, papillary, solid, and lepidic (formerly bronchioloalveolar) patterns

From the image you’ve provided, the presence of glandular structures with cells that resemble goblet cells suggests a significant mucin production. If this is a lung primary, it leans more towards a mucinous adenocarcinoma. However it’s essential to consider the entire context, including clinical history, radiological finding and other histological areas of the tumor. Immunohistochemical staining can also be of great utility in confirming the tissue origin and subtype.

---

Question 3: What about this area? Is this cancer or something else?

Pathrick’s Answer: This image appears to depict a section of epithelial tissue with evident glandular architecture. Key observations include:

1. Pseudostratified Columnar Epithelium: The cells display a columnar morphology and nuclei are positioned at different heights, which is characteristic of pseudostratified epithelium.
2. Ciliated Cells: The apical surface of some cells seems to demonstrate cilia.
3. Presence of Goblet Cells: Interspersed between the columnar cells are cells that resemble goblet cells, characterized by their clear, mucin-filled cytoplasm.
4. Basement Membrane: The epithelial cells are resting on an intact basement membrane.
5. Underlying Stromal Tissue: Beneath the epithelium, there is some stromal tissue with scattered inflammatory cells.

Given these observations, the tissue closely resembles normal respiratory epithelium, possibly from the bronchus or bronchioles, rather than a malignancy. While there are some scattered inflammatory cells in the stroma, which suggest a mild inflammatory response, the epithelial architecture is preserved and does not show overt features of dysplasia or malignancy.

However, as always, it’s crucial to interpret these findings in the context of the patient’s clinical presentation, radiologic findings, and other histologic areas. It’s also important to ensure that representative areas are being evaluated, especially if there’s a clinical or radiologic suspicion of malignancy.

---

Question 4: OK, the mucinous adenocarcinoma we just discussed is in the left upper lobe, I also found a different lesion that looks like in the attached image from the left lower lobe. Do you think it is the same tumor or something different?

Pathrick’s Answer: The image provided showcases cells with a distinct histological appearance from the previous mucinous adenocarcinoma:

1. Cellular Appearance: The cells seem to be more cohesive with an eosinophilic cytoplasm. The nuclear-to-cytoplasmic ratio appears high, and the nuclei look hyperchromatic with some discernible nucleoli.
2. Lack of Mucinous Features: Unlike the previous image, there is an absence of evident mucinous stroma or prominent gland formation. Given these features, this histological appearance suggests a different type of neoplastic process than the mucinous adenocarcinoma discussed previously. It is essential to consider a comprehensive clinical and radiological correlation and potentially utilize immunohistochemical stains to further characterize the lesion and differentiate between primary and secondary tumors or between different subtypes of primary lung tumors.

Matthew: What stains could I use to help in this case?

Pathrick’s Answer: To further characterize the lesion and differentiate between potential primary and secondary tumors or between different subtypes of primary lung tumors, several immunohistochemical (IHC) stains can be employed:

1. TTF-1 (Thyroid Transcription Factor-1): A commonly used marker for primary lung adenocarcinomas. Positive staining supports a lung primary, especially for adenocarcinoma. 
2. Napsin A: Another marker supportive of primary lung adenocarcinoma when positive.
3. CK7 and CK20: These cytokeratins can help in differentiating primary lung tumors from metastatic tumors. Many lung adenocarcinomas are CK7 positive and CK20 negative, whereas certain metastatic tumors (e.g., colorectal adenocarcinoma) may show a CK7 negative/CK20 positive pattern.
4. p40 or p63: These markers can be useful if there’s a suspicion of squamous cell carcinoma. They are typically positive in squamous cell carcinoma and negative in adenocarcinoma.
5. CDX2: A marker for gastrointestinal differentiation. Positive staining would suggest a gastrointestinal origin, such as metastatic colorectal adenocarcinoma.
6. Mucin Stains (e.g., MUC1, MUC2, MUC5AC): These can help in confirming mucinous differentiation if that is a consideration.
7. ALK (Anaplastic Lymphoma Kinase): For identifying ALK rearrangements common in a subset of lung adenocarcinomas.
8. EGFR and KRAS: Molecular testing or IHC can be utilized to identify mutations in these genes, which have therapeutic implications in lung adenocarcinoma.

The choice of IHC stains will be influenced by the differential diagnosis based on the histological appearance, clinical history, and radiological findings. It’s crucial to integrate all these aspects for an accurate diagnosis.

Matthew: Thanks. This new tumor in the left lower lobe is positive for p40 and negative for TTF-1. So, how should I classify the left lower lobe tumor and how should I stage these two tumors in the left upper and lower lobe?

Pathrick’s Answer: The immunohistochemical profile of the left lower lobe tumor being positive for p40 and negative for TTF-1 strongly suggests a squamous differentiation, which is consistent with primary squamous cell carcinoma of the lung.

In terms of classification and staging:

1. Classification:

• Left Upper Lobe Tumor: Based on previous discussion, the left upper lobe lesion is consistent with a mucinous adenocarcinoma.
• Left Lower Lobe Tumor: The tumor is consistent with a primary squamous cell carcinoma based on the IHC profile.

2. Staging:

The staging would be based on the TNM classification system from the American Joint Committee on Cancer (AJCC) and the Union for International Cancer Control (UICC). In the scenario where multiple tumors are identified in a single lung and they are of different histologies, they are considered separate primary tumors. 

• T descriptor: If these two lesions are separate primaries and there is no evidence of direct extension between them, then each tumor will be measured and staged individually. If, for example, both lesions are each 3 cm or less in greatest dimension, they would each be classified as T1. However, the presence of two tumors in the same lobe would make it a T2 (a Tla or Tib plus an additional tumor in the same lobe)
• N descriptor: This would be determined by the status of regional lymph nodes: mediastinal, hilar, and intrapulmonary: Nodes need to be sampled and examined histologically
• M descriptor: This would be determined by the presence or absence of distant metastases. This requires a thorough clinical evaluation, including imaging

Overall, the patient would be staged by combining the highest T, N, and M descriptors. If the tumors have different stages, the higher stage will be used for treatment planning and prognosis. In your case, given that you have two different histological types in the same lung, each would be considered a separate primary and the higher stage of the two would be used for treatment planning.

It is essential to collaborate with a multidisciplinary team, including radiologists and oncologists, to ensure accurate staging and appropriate therapeutic management.

---

Question 5: In another case we found an area like this and the pathologist I am working with said this was lepidic adenocarcinoma. Can you explain to me in simple terms what lepidic adenocarcinoma is? Maybe use an analogy.

Pathrick’s Answer: Lepidic adenocarcinoma is a subtype of lung adenocarcinoma. To put it in simpler terms using an analogy, consider the lung tissue as a garden. In a healthy garden, plants grow in an organized manner, with each plant having its own space. In the context of lepidic adenocarcinoma, imagine a particular type of vine that doesn’t grow deep into the soil, but instead spreads horizontally across the surface. Similarly in lepidic adenocarcinoma, the cancerous cells grow predominantly along the pre-existing structures (like the alveoli, the small air sacs in the lungs) without invading deeply. This pattern of growth is what gives it the term “lepidic” which is derived from the Greek word for scale resembling the spread of scales on a fish.

Therefore, in our analogy, instead of aggressive weeds that dig deep into the garden and displace other plants, lepidic adenocarcinoma is like a surface vine that spreads across without digging deep. This pattern is often less aggressive than other forms of lung cancers. but it’s essential to detect and manage it appropriately,

Matthew: Surrounding the tumor we found some areas that my attending said were organizing pneumonia. Could you explain what we are seeing in the image that makes it organizing pneumonia? Analogies are helpful!

Pathrick’s Answer: Certainly. Organizing pneumonia (previously known as bronchiolitis obliterans organizing pneumonia or BOOP) is a lung condition where there is a particular pattern of lung injury, which is typically related to inflammation and repair processes. Here’s how we can relate the findings on the image to the concept of organizing pneumonia:

1. Masson Bodies: Organizing pneumonia is characterized by the presence of Masson bodies, which are polypoid plugs of loose connective tissue or fibroblasts within alveolar ducts, alveoli, and often bronchioles. In the provided image, we can see some areas with plugs or elongated structures within air spaces that resemble these Masson bodies.
2. Inflammatory Cells: Surrounding these plugs, one might observe an infiltration of inflammatory cells like lymphocytes and macrophages, which is suggestive of an active inflammatory process.
3. Preservation of Lung Architecture: Despite the inflammation and fibrosis, the underlying lung architecture in organizing pneumonia is typically preserved, unlike in other conditions such as usual interstitial pneumonia (UIP) where the architecture is destroyed.
4. Alveolar Walls: The alveolar walls might appear slightly thickened due to the Infiltration of inflammatory cells, fibroblasts, and loose connective tissue.

Imagine a garden (the lung) after a storm. Some pathways (bronchioles and alveoli) are blocked by debris (the Masson bodies or plugs of fibroblasts), gardeners (inflammatory cells) are busy trying to clean up and repair the damage. Even with the mess and the ongoing repair, you can still recognize the original layout and structure of the garden (preservation of lung architecture).

This analogy simplifies the process, but it might help to visualize the repair and inflammation mechanisms occurring in the tissue. Always correlate with clinical and radiological findings for a comprehensive diagnosis.

---

Grading my student

So, what are my thoughts on Pathrick’s performance? Did he make the grade?

In these experiments I have collaborated with Pathrick in the capacity of a resident trainee, tackling simulated case scenarios and posing specific questions relevant to each case. The images utilized were sourced from publicly available whole slide images (WSI). Although these images are unique, there remains the possibility that other images from the same slides could have been incorporated into ChatGPT's training data. A valuable metric for evaluating the model's performance is comparing with human benchmarks. In my assessment, the quality of responses from ChatGPT would surpass what is typically expected from an intermediate to advanced pathology resident. One of the most useful features of these models, in my opinion, is their ability to generate novel analogies for describing various pathological processes. For instance, I found the analogies of describing lepidic cancer as ‘vines’ and organizing pneumonia as ‘post-storm debris’ to be exceptional. These analogies allow learners to conceptualize complex processes within a familiar mental framework, and I intend to incorporate them into my own teaching going forward.

However, the model's outputs do tend to be verbose, which is a recognized characteristic of ChatGPT. Upon a closer examination of the images, it becomes apparent that certain features, such as goblet cells in the airways, are either absent or not discernible. This highlights a crucial aspect regarding the functioning of ChatGPT and other transformer-based models: they are fundamentally predictive engines. Employing complex mathematical algorithms, these models attempt to anticipate the desired response, a mechanism that can sometimes result in data hallucinations. In this particular instance, the hallucination is subtle; the model identifies an airway structure and subsequently describes features commonly found in airways, such as goblet cells, despite the absence of visual confirmation.

The potential for applications of such models is significant, yet the issue of data hallucination remains a major challenge that requires attention. As it stands with the present generation of these tools, it is critical that their outputs undergo an assessment by experts to ensure the accuracy of the generated responses.