To Look Beyond what eyes can see! — “Radiomics” for Lung Cancer
Here’s a quick explainer regarding our latest work published in Lancet Digital Health. The entire article can be found here.
Cancer is a leading cause of death worldwide. As per the world health organization (WHO), today, one in the six deaths is due to cancer and in total accounts for ~ 9.6 million deaths annually. The impact is worse in low and middle-income countries, and this number is only expected to increase.
Amongst all the cancers, lung cancer is considered as the foremost contributor to cancer-related mortalities, not only because it’s the most common one amongst both men and women, but also because it is relatively difficult to diagnose at an early stage. Cancer treatments provided at an early stage usually has a high impact on improved survival. For example, 89% of women diagnosed with breast cancer survive longer than 5 years with effective early-stage treatments. Whereas for lung cancer patients, the five-year survival rate has remained 18% — a statistic that has not changed significantly in decades.
“Early-Stage” lung cancer accounts for stage I and II Non-Small Cell Lung Cancer. For these cancer cases, tumor size usually appears less than 4cm on the radiographic scan, and cancer is contained only in the primary location (not metastasized). With an increase in computed tomography (CT) technology, nowadays, it has become relatively more accessible and apparent to detect these small nodules at a relatively early stage. What is a CT scan? A CT scan combines a series of x-ray images taken from different angles around the patient’s body and uses computer processing techniques to create cross-sectional images (slices) from inside your body. It produces 2-dimensional images of a section of the body (in this case, lung), and the combined data is used for constructing the 3-dimensional lung images.
Looking at the treatments of these early-stage cases, one of the best standard options includes providing chemotherapy following surgery to avoid potential tumor reappearance or recurrence. Even though all these early-stage cases are eligible for adjuvant cytotoxic chemotherapy, more than 50% of these patients may have low-risk disease and hence may not receive added benefit from it, while suffering its side-effects. From an economic standpoint, unnecessary adjuvant chemotherapy for early-stage NSCLC results in a loss of over ~$35000 for each quality-adjusted life-year loss. The current decision whether or not to give chemotherapy to each patient is based on clinical prognostic factors, but there’s no definite validated CDx test or biomarker to predict added benefit of adjuvant chemotherapy for early-stage NSCLC directly.
Artificial Intelligence in Clinical Applications
Artificial Intelligence (AI) includes computer science algorithms, machine learning, deep learning, pattern recognition, computer vision techniques, all packed up into one term. AI computes approximate conclusions to the problem without direct human involvement. It can identify meaningful relationships in the raw data, which would be difficult — or almost impossible — for humans to solve or predict.
In clinical settings, AI and computerized decisions can be integrated into routine clinical practice for diagnosis, prognosis, and predicting individualized treatment response in various clinical problems. Although there has been notable progress in the medical fields regarding cancer treatments as well as a significant improvement in AI models in recent years, there is a need to develop an integrated clinical research environment to implement AI in the routine clinical practices. In our work, we have developed an imaging-based AI biomarker that predicts survival as well as the added benefit of adjuvant chemotherapy for each early-stage lung cancer cases. Furthermore, we have constructed a combined model, including clinical factors that are used in routine practice to our developed imaging-signature to predict response to adjuvant chemotherapy. The combined model outperformed identifying patients that would benefit from adjuvant-chemotherapy compared to the traditional clinicopathological factors used in the clinical practice.
What is “RADIOMICS”? — Images are more than pictures, they are data!
Radiomics is an emerging translational field of research focused on extracting high-dimensional data from clinical, radiographic images and finding its association with clinical output. The concept underlying the Radiomics is that these clinical images contain data, which reflect the underlying pathophysiology of tissue, which is not apparent to human eyes. Radiomics focuses on extracting different kinds of textural features from a defined region of interest (ROI) on the clinical scans. The ROI can be any annotated region on the clinical scans- it can be a tumor region, the area immediately outside the tumor, or even the normal tissue appearing on the scan. The texture extracted from the defined ROI can exhibit different levels of complexities. They can express properties regarding shape and appearance, voxel intensities as well as the spatial arrangement of the intensity values at the voxel level as observed on the imaging scans. Radiomics refers to using these large numbers of parameters extracted from single ROIs. These parameters are mathematically computed and processed with advanced statistical methods under the hypothesis that an appropriate combination of them, along with clinical data, can express significant tissue properties, useful for diagnosis, prognosis, or treatment in an individual patient. Radiomics is often further combined with machine learning models to learn to detect hard-to-discern patterns from large and complex data sets and make decisions on the unseen, new data.
Overview of our Main Results
We used Radiomics features extracted from the tumor region and immediately outside the tumor region (15mm outside the tumor) from the CT scans of early-stage lung cancer cases to develop an imaging-based score to predict the survival of these patients.
We observed that the developed imaging score had prognostic value, i.e., these textural patterns were significantly different between aggressive cancer nodules compared to the not-so-aggressive cancer nodules. We divided the developed score into three groups-high, intermediate, and low risk.
The patients having a high score were observed to have better survival when they received chemotherapy following surgery. Whereas, patients with a low score were observed to be doing well with surgery alone.
Further, we integrated the developed score with clinical factors that are usually used for decision making regarding chemotherapy. The integrated model included developed imaging scores with pathologic T-, N- Stage, and lymphovascular status. We calculated the estimated survival benefit for all the patients using the clinico-radiomics model.
When the estimated survival benefit was less than 20%, the patients who received chemotherapy had better survival as compared to patients who underwent surgery alone.
To analyze what these Radiomic patterns mean on a biological basis, we investigated the correlation of prognostic Radiomics from CT scan to histological tissue scans as well as mRNA sequencing data for the sub-cohort of these patients.
We observed that these prognostic Radiomic patterns were correlated with the interplay between cancerous cell clusters and lymphocyte cell clusters observed on the whole slide tissue scans, a biomarker to identify aggressive lung cancer patients. On a genomic analysis, the developed score was observed to correlate with angiogenesis, proliferation, cellular differentiation, T-Cell, and Lymphocyte activation, etc. These are biological pathways observed during cancer development and progression.
In conclusion, we constructed a Risk Score (QuRiS) and combined Clinico-Radiomics model (QuRNom) that potentially would improve identifying ideal candidates for adjuvant chemotherapy and also would spare the patients with unnecessary overtreatment and adverse effects of chemotherapy. Going forward, integrating Radiomics into routine clinical practice after prospective validation of these models would potentially be the final goal of our study. Considering that the developed model only uses CT scans, which are routinely acquired in the clinical practice for all lung cancer cases, integrating this model would not disrupt the overall clinical workflow. In fact, it would give an added prognostic biomarker for oncologists, which they can take into consideration while making the final decision.