Lung cancer is the leading cause of cancer-related death in men and women. When a doctor suspects a patient has lung cancer, he or she will perform a thorough history and physical examination, order imaging tests to identify any masses, and order biopsies to examine suspicious lung tissue under the microscope.
Pathologist Diagnosis of Lung Cancer
The pathologist analysis of lung tissue biopsy under the microscope is the cornerstone of cancer diagnosis, staging, and treatment planning.
The three main types of lung cancer are adenocarcinoma, squamous cell carcinoma, and small cell carcinoma, which look like the following under a microscope:
Classically, biopsy slides are analyzed using TNM staging, which is an international standard classification to assess tumor size (T), spread of cancer to local lymph nodes (N), and metastasis or spread to distant parts of the body (M). Pathologists also describe different morphologic features of the cancer. For example, do the malignant cells resemble mucus-secreting glands (adenocarcinoma)? Or do the cells look like thin, rectangular cells like the ones found on the surface of your skin (squamous cell carcinoma)? Nowadays pathologists also apply special stains to classify the chemical characteristics of cancer cells and describe the genetic drivers that are causing the cancer to grow. These pathology reports help oncologists determine the appropriate therapy that works best for each patient.
A pathology report for a biopsy for a suspected lung cancer can look like the following:
AI analysis of pathology slides require high-resolution, whole-slide imaging to view the entire tissue sample on the slide.
However, nowadays most pathologists analyze tissue samples with glass slides under a microscope. One significant concern is that there is significant variability and inconsistency between different pathologists in interpretation. Several studies have shown that the agreement between pathologists is low (K= 0.41-0.69), regardless if they are community pathologists or pulmonary pathology experts (Stang et al, Lung Cancer 2006, Grilley-Olson et al, Arch Pathol Lab 2013, Thunnissen et al, Mod Pathol 2012). This variability may lead to poor therapeutic choices in patients, and is part of the reason why pathology analysis is often “re-done” with a second opinion when a patient goes to a different hospital. Second opinions involve the annoying steps of storing glass slides in boxes, shipping them by mail to other specialists, and hoping no material becomes broken or lost in the process when waiting for the diagnosis.
Digital pathology can more easily help in rapid quality, safety, and communication between different caregivers. Increased adoption of digital pathology has already started. The first digital pathology solution cleared for primary diagnosis was FDA cleared in 2017 by Philips (https://www.usa.philips.com/healthcare/solutions/pathology). The Philips IntelliSite Pathology Solution is a high-throughput automated scanner that can scan up to 300 glass slides at a time for 40x image magnification. This technology is super promising and is currently being trialled at academic health centers across the globe.
AI in Lung Cancer Diagnosis
Two key recent studies demonstrated the promise of artificial intelligence in lung cancer diagnosis. Links here:
Yu et al. sought to create a machine learning algorithm to better predict lung cancer patient prognosis. To accomplish this study, they analyzed 2,186 images from the Cancer Genome Atlas and 294 images from Stanford Tissue Microarray Database, extracted 9,879 quantitative image features, and used machine learning to identify the features that best distinguished short term survivors from long term survivors. They then showed that imaging features identified by machine learning better characterized patient survival than clinical tumor stage or grade.
Coudray et al. sought to differentiate between adenocarcinoma, squamous cell carcinoma, and normal tissue using deep learning. To accomplish this, they also relied on whole-slide images from the Cancer Genome Atlas. In this study, their deep learning algorithm was able to identify different subtypes of lung cancer with a high accuracy (AUC of 0.97), similar to pathologist manual annotation. Not only that, they were able to show that their AI algorithm could predict at least 6 genetic drivers of cancer from image alone with high sensitivity and specificity.
Promising AI Companies in the Pathology Space
Two promising companies I am following in the pathology AI space are PathAI and Proscia.
PathAI (https://www.pathai.com) is a Cambridge based company that recently raised $11 million in Series A funding. They recently won the Camelyon challenge that identified metastatic cancer in lymph nodes with an error of 0.6 percent (better than pathologist error of 3.5 percent). They have recently partnered with Bristol-Meyers Squibb and Philips, and just last week they received ISO13485 certification.
Proscia (https://proscia.com/) is a Philadephia based AI company founded by David West, a bright, ambitious man I was fortunate to meet when I was an undergraduate at Johns Hopkins. Proscia has raised approximately $12.3 million and recently released an update to Concentriq, a digital pathology platform for whole-slide imaging analysis for clinic, research, and pharmaceutical companies.