A Feeling Out Process
By Pamela Q. Fernandes, MD
Radiology Today
Vol. 20 No. 3 P. 16

Texture imaging is moving cancer assessment forward.

In November 2018, Memorial Sloan Kettering Cancer Center (MSKCC) hosted the Pancreas Cancer Symposium. For a disease with a grim prognosis and lack of research breakthroughs, alternative assessments are desperately sought. One alternative proposed at the conference was to improve prognostication as it relates to pancreatic cancer. To that end, Richard Kinh Gian Do, MD, PhD, a radiologist at MSKCC whose expertise is in imaging for liver cancer, bile duct cancer, and pancreatic cancer, talked about tumor texture analysis (TA) and texture imaging.

Do expresses enthusiasm about what appears to be a promising area of body imaging. “Texture analysis is a subset of computer-based image analysis where we look for patterns in a quantitative way,” Do says. “It is the interrelation of the pixels on the grey scale from black to white.”

Mathematical equations are applied to the measured greyscale pixels, and features are generated. These features and patterns are analyzed to produce quantitative results or outcomes. For radiologists, pattern recognition is second nature. After all, it is, as Do says, “what our brains are trained to do.” But the struggle is with subtle gradations.

“Black and white, yes and no are easier,” Do says. “Accuracy, sensitivity, and specificity are important, but, when it comes to tumor biology, there’s no black and white; there’s a spectrum of the disease, even in resectable pancreas cancer.”

Though he hopes a cure may emerge, he observes that current treatments have a 20% survival rate. So, with texture imaging, patients can be directed to a better treatment paradigm with more predictable tools. For example, with a predictive tool such as TA, physicians can tell patients to opt for treatment but inform them that they’d benefit more from up-front chemotherapy based on the TA results. After six months, if they have no serious complications, surgery could be suggested.

There’s always debate among the treatment teams about which patients are better candidates for neoadjuvant therapy vs surgery first. With texture imaging, physicians can assess outcomes by looking at thousands of images quantitatively, predicting who the better candidates for surgery are, and determining which therapy is best for each patient.

Satellite Images
Texture imaging has been around for more than a decade and has been used extensively in the industrial and defense fields. What makes TA potentially useful for radiology is the high-quality medical imaging data that have been accumulated during the past decade, which can now be mined for trends and patterns, Do says. Do’s research team worked with a computer scientist to determine the most useful areas of focus and decided to begin with liver imaging for liver failure, followed by colorectal metastases, pancreatic tumors, and hepatocellular carcinomas.

“Our studies came from having a conversation in a multidisciplinary setting with hepatobiliary surgeons about predicting grades of different tumors,” Do says. “It’s the same technique across all the cancers because we want precision medicine, knowing that cancers are all heterogeneous.”

One of the challenges of texture imaging is that there is a subjective element involved; two radiologists may view texture quite differently. TA is basically the use of equations to compare adjacent pixels, but Do says different research groups apply TA in different ways. He uses the example of comparing a checkerboard to a black and white cookie; each one may have the same amount of black and white, but equations can identify and describe the differences in distribution. For this reason, standardized equations are necessary.

Part of the difficulty lies in the fact that images come to radiologists processed, filtered, and smoothed by the proprietary methods of the imaging equipment, Do says. In addition, each researcher or radiologist is able to filter his or her own images, which can make a significant difference from researcher to researcher, he adds. Although TA identifies signal patterns within the images, collaboration is needed to smooth out the variability.

“We are at the stage of hypothesis generations of correlative studies to look for these patterns,” Do says. “One of the criticisms is that there’s no biological basis for this as compared to genetic analysis. But just like genetics, there was a lot of noise in the early days because there was all this data and we didn’t know how to make sense of it.”

The technology is in its infancy, Do says, and, although finding a biological basis for texture signals is important, it’s not 100% necessary. He likens the process to weather prediction; barometric pressure and wind speed are important variables, but these emerging properties require a bigger picture: a satellite map.

“You need to see the hurricane that is an ocean away. With imaging, we are the satellite imagers of cancers,” Do says. “There are emerging properties that only CT can detect and a computer can decode that, without the big picture, you won’t be able to predict.” What has intrigued Do is that there is an opportunity with TA to make bold predictions about cancer that can’t be achieved with blood testing.

Clinical Use
TA can be applied to any modality, but the best choice varies by anatomic region. Do says CT is best for hepatobiliary and lung imaging because there are more data available, while MRI is preferable for the prostate and brain. Breast imaging can be served by a combination of mammography and MRI. However, as Do points out, MRI has artifacts that humans are able to recognize, while a computer would need to be trained to spot them. Ultimately, the most important consideration is where the data are.

“The bigger the data set, the more reliable the prediction will become,” Do says.

The results so far have been promising. Do and his team published a study on colorectal cancer patients showing that TA was able to identify patients who were likely to develop liver metastases—before the metastases were visible to the human eye.

“Is it because there are already metastases there that are altering the blood flow in some subtle ways that the CT is detecting early on?” Do asks. “Or are there underlying biological differences in the liver that are priming it for liver metastases to develop, such as micrometastatic niches?”

The researchers at MSKCC are actively researching all of the possibilities. Do is optimistic for similar results in pancreatic cancer as well. He is especially excited about the prospect of finding a quantitative answer.

“We all have a gestalt sense of when a patient looks worse than what the imaging shows or what’s communicated in a report, but, as radiologists, we’re trained for ‘yes or no’ questions,” Do says. “We don’t have the ability to have a sliding scale of aggressiveness of a tumor. And being able to add that to the equation can change the conversation with the patient in the long run.”

The analogy he makes is, if you want to buy an apartment, you want to know whether the interest rates are going to go up or down. Most people will wonder whether they should wait. If someone could tell them that there is a 90% chance rates will be lower, they would wait. With chemotherapy or surgery, the question is always, “Should I do chemotherapy or surgery up front?” For pancreatic cancer, being able to determine that there’s a 90% chance surgery will not be helpful could be useful for the discussion between the medical team and the patient.

Prognostic Potential
Currently, surgery is the frontline option for pancreatic cancer. But there’s no way to know whether it will be successful until physicians learn the lymph node status, tumor grade, whether the resection was clean, or whether there was a remnant left behind, Do says.

“But if you can have a reliable predictor, a nomogram that’s clinically approved, that can change the conversation for those who are not the best surgical candidates or for those interested in chemotherapy,” he adds.

In his own studies from 200 patients, Do’s team has shown that TA can have the same predictive value as the postsurgical standard of care model, the Brennan score, for predicting pancreatic cancer survival. He doesn’t don’t want to overpromise, however.

“The biomarker needs to be validated by multicenter evaluation studies,” Do says. “We want to test our model in blinded data sets from other institutes and say that, under different conditions, we could replicate the same results.”

Do cautions that single center data are not enough evidence to make bold predictions. He says the challenge is to interpret the results that are being published.

“We have to temper the excitement because of the lack of methodology,” Do says. “It’s still early days. People are all applying it in different ways; there’s no right answer yet. We’ll converge to the right methodology in the next few years, and then we can compare results across centers.”

Getting Started
If radiologists are interested, they can get started on TA right now, according to Do. He notes that there is a significant amount of literature available; there is also open source software called Pyradiomics.

“As long as you have a tumor that’s segmented, the tools are free and you can apply them,” Do says. “The biggest challenge is to interpret the results. That’s where you need statisticians because, if you don’t know what you’re doing, the software will spin out a lot of data and you wouldn’t know how to test for significance.”

TA requires correction for the hundreds of thousands of features that it generates, Do says. If correction is not done properly, TA will find abnormalities by chance.

“Unless you’re aware of these pitfalls, you may publish successfully, but it’s not adding to the literature,” Do says.

Successful application of TA requires multicentric trials to validate the data. Do says TA for lung cancer is the closest to clinical use because there are a great deal of data associated with it, and there have been successful multicentric TA trials. He says there is also significant potential to adapt TA to breast cancer evaluation over the next decade, again because of large databases associated with the disease.

“For rarer cancers like the pancreas, it will take more multicentric studies to be brought to clinical care,” Do says.

That’s something Do is passionate about. He’s hoping to generate a collaborative spirit to make it happen because, without multicentric collaboration, he feels it will be very difficult to bridge the distance to clinical care. To do that, Do believes it is best to have a central repository where people from multiple disciplines can share data. He notes that the ACR Data Science Institute was created to serve this purpose.

“The best idea may not come from radiologists or even a computer scientist in the field; it could come from someone outside the medical community. This data would be multicentric, and it would give patients the best chance in the long run, faster,” Do says. “The ACR is committed to data privacy concerns and governance. I encourage people to engage and share their data with that repository. It benefits all patients to have a single authority to say this is the best data.”

Do says TA is the tip of the iceberg when it comes to cancer imaging. Big Data companies have built repositories of billions of annotated data sets and dedicated a significant amount of funding toward analyzing those data sets. With the same resources and a collaborative spirit, Do strongly believes that rare cancers such as pancreatic cancer can benefit from a quantitative tool such as TA. It’s a daunting and challenging task, but he is optimistic for the future.

“Texture imaging is a subset of radiomics. With radiomics, there are opportunities for machine learning and neural networks. In the long term, neural networks could even beat texture analysis because [TA tools] were tools applied to other industries that were then applied to medical imaging but were not designed primarily for medical imaging,” Do says. “Right now, we can do [TA] because the data sets are smaller but, in the long run, once we have more data, [neural networks] would be a game changer. With tens of thousands of cases, machine learning will beat texture analysis.”

— Pamela Q. Fernandes, MD, is a doctor, author, and medical writer who specializes in new breakthroughs in medicine. More information is available at www.pamelaqfernandes.com.