Deep Learning in NSCLC Care

From MedPage Today

NSCLC Care informed by Deep Learning is a new application of machine learning.

In a 2018 JAMA Viewpoint, experts noted that the most successful applications of deep learning in medicine have been for analyzing medical images and that the “potential of deep learning to disentangle complex, subtle discriminative patterns in images suggests that these techniques may be useful in other areas of medicine.”

Four years later, researchers in Canada are advancing the role of machine learning in lung cancer care via a “fully automated imaging-based prognostication technique [IPRO]” as applied to pre-treatment CT imaging. The IPRO uses “deep learning to predict 1-year, 2-year, and 5-year mortality from pretreatment CTs of patients with stage I-IV lung cancer,” explained Felipe Soares Torres, MD, PhD, of the University of Toronto, and colleagues.

They explained in JCO Clinical Cancer Informatics that the current CT analysis “relies upon manual localization, classification, and measurement of lesions and is subject to interobserver and intraobserver variability,” but that “a fully automated approach, in which a system would analyze the entire thorax in a CT, may complement traditional TNM [tumor-nodes-metastases] staging of patients with lung cancer and provide greater prognostic power in a standardized manner that does not require manual steps subject to interobserver variability.”

In their retrospective study, the researchers tested IPRO in six datasets from the Cancer Imaging Archive, with pre-treatment CTs of 1,689 patients, the majority of whom had non-small-cell lung cancer (NSCLC) and had available TNM staging information.

The team compared the association of IPRO and TNM staging with patients’ survival status and assessed an Ensemble risk score that combined the two. Highlights from the study included the following:

  • IPRO showed similar prognostic power versus TNM staging in predicting 1-year, 2-year, and 5-year mortality
  • The Ensemble risk score yielded superior performance across all time points
  • IPRO stratified patients within TNM stages, discriminating between highest- and lowest-risk quintiles, such that adenocarcinoma was the most common histologic subtype in patients in the lowest-risk decile, followed by squamous cell carcinoma and large cell carcinoma (highest risk). Those with TNM stages I and II accounted for more than three-fourths of patients in the lowest-risk decile but less than 30% of patients in the highest-risk decile

“Subject to further prospective validation, IPRO could have the potential to guide treatment intensification and deescalation strategies,” Torres and colleagues said.

In this interview, Torres, assistant professor of radiology in the Joint Department of Medical Imaging, discussed some of the details of the study, and how the IPRO could potentially be deployed in practice.

Professor of Radiology on NSCLC Care Informed by Deep Learning

Is IPRO an example of radiomics or is it a different application of deep learning?

Torres: Traditional radiomics involves manually segmenting regions of interest (e.g., tumors) and extracting selected features (e.g., texture, etc.) from those regions. Our deep-learning approach, called IPRO (Imaging-based PROgnostication), enables the model to take into account any potentially relevant imaging features from the entire thorax to predict mortality — not just those features that are currently perceived as relevant or quantifiable.

The paper states: “In our model development, we enabled IPRO to consider prognostic features across the entire thorax, eliminating the need for radiologists to manually annotate regions of interest, such as primary tumors. Manual annotation is a time-consuming process, requires radiologic expertise, is subject to inter-reader variability, and enforces the assumption that only annotated regions of interest are correlated with outcomes.” Does this feature make IPRO unique among the various deep learning models currently under investigation?

Torres: While most deep learning models are trained to replicate radiological tasks like abnormality classification, organ segmentation, or lesion detection and measurement, IPRO in contrast is a prognostic model that predicts mortality risk without being explicitly “told” to replicate radiological guidelines (e.g., TNM staging).

How would you respond to someone interpreting this as IPRO simply eliminating the need for a radiologist altogether?

Torres: Our paper illustrates the benefit of combining IPRO and TNM staging to achieve superior prognostic accuracy, which has potential to inform more personalized treatment decisions. TNM staging informs an oncologist about the anatomic extent of disease and guides the selection of appropriate treatment.

In contrast, IPRO’s mortality risk score alone, while accurate, does not yet convey the origin of risk — whether the high mortality risk due to the large tumor, old age, or cardiovascular comorbidity — which is needed to make treatment decisions.

Model explainability is thus an important aspect of our current and future work, and why we sought to measure IPRO’s attention using heatmaps and association with known prognostic factors. The model is intended to provide important and additional information for a more personalized treatment decision, not to replace or eliminate radiologists.

Feedback on Value of Data in Deep Learning Cancer Care

Have you received feedback (anecdotal or otherwise) from your oncology/radiation oncology/thoracic specialty colleagues on the clinical value of the data generated by IPRO?

Torres: There is interest from oncology colleagues in exploring how IPRO could enhance treatment decision-making (e.g., treatment intensification for high-risk patients and deescalation for low-risk patients), so improving IPRO’s explainability is an important focus of future work.

We are also exploring how IPRO’s ability to comprehensively stratify patients could improve clinical trial design via better baseline randomization, and, when applied to follow-up scans, improve the association between RECIST-based surrogate endpoints and overall survival. Further, we plan to incorporate IPRO as an exploratory endpoint in prospective studies and explore combinations of IPRO and other biomarkers, such as circulating tumor DNA.

In a 2020 JCO Oncology Practice paper, researchers described a “fast-and-frugal” decision tree for lung cancer management. Could IPRO potentially play a role in or with these types of decision-making algorithms?

Torres: IPRO could be included as a cue in the decision trees, since it is obtained from pre-treatment CT images, an integral part of diagnosis and staging of patients with lung cancer. One important aspect related to evidence-based medicine is the transparency of the decision-making process, particularly transparency regarding the basis for articulating the decision. We are currently exploring ways to better understand the mortality prediction provided by IPRO. This will be crucial to integrate IPRO in the decision-making algorithms for treating NSCLC.

A 2021 JCO Oncology Practice paper discussed the importance of clinical pathways in promoting consistent evidence-based care “in an increasingly dynamic [NSCLC] treatment landscape.” Could the IPRO model play a role in clinical pathways and the promotion of evidence/value-based care?

Torres: We demonstrated that IPRO may provide additional prognostic insight based on both known and unknown features present in pre-treatment CTs. Incorporating such information into clinical pathways may help in selecting the right treatment for a particular patient. Additional studies, however, will be required to define the contribution of IPRO in the management of patients with NSCLC.

We are exploring the use of IPRO for measuring treatment effect in clinical trials — thereby in an environment where evidence-based conclusions can be provided. But it is yet to be explored how clinicians would use IPRO in clinical practice.

The paper states that it is yet to be determined how IPRO will perform when additional clinical variables like pulmonary function and treatment type are introduced into the model, and that future work will explore how clinical variables and longitudinal imaging can be incorporated to further refine prognostic accuracy. Has your group started this research?

Torres: Yes, we are excited to publish additional work over the coming months. For example, we are currently exploring how information from pulmonary function tests can be predicted from the CT images using deep learning — information that can be extremely useful in expediting treatment decisions for patients with lung cancer. We are also open to collaborating with additional sites/oncologists interested in exploring various applications of IPRO, including expanding it into other tumor types and pathologies.

Read the study here.

The study was supported by Altis Labs. Two co-authors are company employees.

Torres disclosed a relationships with Altis Labs. Co-authors disclosed relationships with Altis Labs, AstraZeneca, AbbVie, Verity Pharmaceuticals, Sanofi, Varian Medical Systems, Amgen, Bristol Myers Squibb, MSD, Novartis, Sanofi Genzyme, Takeda, EMD Serono, GlaxoSmithKline, Puma Biotechnology Array, Bayer, Eli Lilly, Guardant Health, Inivata, Novartis, Pfizer, and Roche.

Primary Source

JCO Clinical Cancer Informatics

Source Reference: Torres FS, et al “End-to-end non-small-cell lung cancer prognostication using deep learning applied to pretreatment computed tomography” JCO Clin Cancer Inform 2021; 5: 1141-1150.