Scientists from the Case Western Reserve University digital imaging lab, already pioneering the use of Artificial Intelligence (AI) to predict whether chemotherapy will be successful, can now determine which lung-cancer patients will benefit from expensive immunotherapy.
And, once again, they’re doing it by teaching a computer to find previously unseen changes in patterns in CT scans taken when the lung cancer is first diagnosed compared to scans taken after the first 2-3 cycles of immunotherapy treatment.
And, as with previous work, those changes have been discovered both inside—and outside—the tumor, a signature of the lab’s recent research.
“This is no flash in the pan – this research really seems to be reflecting something about the very biology of the disease, about which is the more aggressive phenotype, and that’s information oncologists do not currently have,” said Anant Madabhushi, whose Center for Computational Imaging and Personalized Diagnostics (CCIPD) has become a global leader in the detection, diagnosis and characterization of various cancers and other diseases by meshing medical imaging, machine learning and AI.
Currently, only about 20% of all cancer patients will actually benefit from immunotherapy, a treatment that differs from chemotherapy in that it uses drugs to help your immune system fight cancer, while chemotherapy uses drugs to directly kill cancer cells, according to the National Cancer Institute.
Madabhushi said the recent work by his lab would help oncologists know which patients would actually benefit from the therapy, and who would not.
“Even though immunotherapy has changed the entire ecosystem of cancer, it also remains extremely expensive – about $200,000 per patient, per year,” Madabhushi said.
“That’s part of the financial toxicity that comes along with cancer and results in about 42% of all new diagnosed cancer patients losing their life savings within a year of diagnosis.”
Having a tool based on the research being done now by his lab would go a long way toward “doing a better job of matching up which patients will respond to immunotherapy instead of throwing $800,000 down the drain,” he added, referencing the four patients out of five who will not benefit, multiplied by annual estimated cost.
New research published
The new research, led by co-authors Mohammadhadi Khorrami and Prateek Prasanna, along with Madabhushi and 10 other collaborators from six different institutions (see list, below) was published this month in the journal Cancer Immunology Research.
Khorrami, a graduate student working at the CCIPD, said one of the more significant advances in the research was the ability of the computer program to note the changes in texture, volume and shape of a given lesion, not just its size.
“This is important because when a doctor decides based on CT images alone whether a patient has responded to therapy, it is often based on the size of the lesion,” Khorrami said. “We have found that textural change is a better predictor of whether the therapy is working.
“Sometimes, for example, the nodule may appear larger after therapy because of another reason, say a broken vessel inside the tumor—but the therapy is actually working. Now, we have a way of knowing that.”
Prasanna, a postdoctoral research associate in Madabhushi’s lab, said the study also showed that the results were consistent across scans of patients treated at two different sites and with three different types of immunotherapy agents.
“This is a demonstration of the fundamental value of the program, that our machine-learning model could predict response in patients treated with different immune checkpoint inhibitors,” he said. “We are dealing with a fundamental biological principal.”
Prasanna said the initial study used CT scans from 50 patients to train the computer and create a mathematical algorithm to identify the changes in the lesion.
He said the next step will be to test the program on cases obtained from other sites and across different immunotherapy agents. This research recently won an ASCO 2019 Conquer Cancer Foundation Merit Award.
Additionally, Madabhushi said, researchers were able show that the patterns on the CT scans which were most associated with a positive response to treatment and with overall patient survival were also later found to be closely associated with the arrangement of immune cells on the original diagnostic biopsies of those patients.
This suggests that those CT scans actually appear to capturing the immune response elicited by the tumors against the invasion of the cancer—and that the ones with the strongest immune response were showing the most significant textural change and most importantly, would best respond to the immunotherapy, he said.
Madabhushi established the CCIPD at Case Western Reserve in 2012. The lab now includes nearly 60 researchers.
Some of the lab’s most recent work, in collaboration with New York University and Yale University, has used AI to predict which lung cancer patients would benefit from adjuvant chemotherapy based on tissue-slide images. That advancement was named by Prevention Magazine as one of the top 10 medical breakthroughs of 2018.
Artificial intelligence (AI), a field of computer science whose origins can be dated back to the 1940s,1 aims to develop computational systems with advanced analytical or predictive capabilities. Such systems are typically designed to solve complex data-intensive problems that require the prediction of, or reasoning about, their underlying phenomena. Machine learning (ML) represents the most successful branch of AI. ML is concerned with the development of programs with the capacity to learn from data.
Such a learning capacity is achieved by incrementally improving on a prediction task based on problem-specific measurements of performance.2 As defined by Tom Mitchell more formally: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E”.3
ML has a solid history of applications in biomedical research,4–6 and is becoming a driver of pre-clinical and clinical oncology research.7,8 Over the past few years, ML’s potential in precision oncology has become more apparent by the reporting of major advances in deep learning (DL), and its application to a variety of diagnostic, prognostic, and other predictive tasks.5,9–11
DL is a sub-branch of ML that comprises a diverse family of computational models consisting of many (deep) data processing layers for automated feature extraction and pattern recognition in large datasets, e.g., multilayer feed-forward neural networks.12,13 DL’s advances in cancer research have been made possible not only by the availability of larger datasets and accelerated computing capabilities, but also by developments in statistical learning theory, algorithms, and open-source software accumulated over the past 4 decades.13
The current visibility of DL in precision oncology has been in large part due to its impressive performance in classifying imaging data in different clinical domains. Examples include: the detection and classification of skin lesions,11 the identification and categorization of lung cancers,14 and the detection of metastases in women with breast cancer,15 all of which apply different versions of a DL technique known as convolutional neural networks (CNN).12,16
Precision oncology research also benefits from a variety of alternative ML approaches for supervised and unsupervised pattern analysis in datasets originating from multiple sources, including tumor-derived omic profiles. This includes, for example, the prediction of oncogenes and tumor suppressors with random forests.17
Key examples of (non-DL) ML techniques include: probabilistic models, kernel-based models (e.g., support vector machines), and decision tree-based models (e.g., random forests and gradient boosting machines; GBM).3,18,19 These and other approaches have provided the basis for promising predictive modelling applications in oncology research.20,21 More detailed examples are provided in the following sections.
To date, ML has played a prominent role in facilitating novel applications that mainly rely on the supervised identification, correlation, and classification of complex data patterns for patient stratification. However, to deliver on the promise of a more precise prevention, detection, and treatment of cancers, other clinically-oriented computational modelling challenges must be tackled. This perspective underlines a selection of such research challenges or requirements for moving the field forward.
It argues that ML offers, yet to be fully tapped, opportunities for enabling precision oncology far beyond relatively well-known applications, such as the supervised classification of single-source omics or imaging datasets. Moreover, there is a need for modelling approaches that can assist researchers and clinicians in better understanding biological causality. To advance and accelerate precision oncology research, the scope of questions and applications that AI can address ought to be considerably expanded (Fig. (Fig.11).
This article discusses four key challenges in ML for precision oncology: dealing with multiple data modalities, insufficient data, interpretability and explanation, and alternative learning approaches. Each section begins with overviews of significant progress achieved to date. The article concludes with an outlook, outstanding questions, and final remarks.
More information: Mohammadhadi Khorrami et al, Changes in CT radiomic features associated with lymphocyte distribution predict overall survival and response to immunotherapy in non-small cell lung cancer, Cancer Immunology Research (2019). DOI: 10.1158/2326-6066.CIR-19-0476
Provided by Case Western Reserve University