Artificial Intelligence Community for Imaging in the Works

Ingrid Hein

November 29, 2017

An open-source artificial intelligence tool could standardize the radiomics algorithms used in clinical trials, make deep learning more accessible to all radiologists, and ensure that trial results can be more easily compared, researchers report.

"A lot of people are doing deep-learning analysis all over the world in isolation," said Hugo Aerts, PhD, from the Dana-Farber Cancer Institute, Brigham and Women’s Hospital, and Harvard Medical School in Boston.

With every lab doing things its own way, it's difficult to compare trial data. But "if researchers start using the same features, names, and algorithms, we can start reproducing and comparing results," he told Medscape Medical News.

Dr Aerts and his colleagues launched to automate the quantification of radiographic phenotypes. The open-source platform extracts radiomics features from 2D and 3D images and binary masks, and helps with the loading and preprocessing of images, such as resampling and cropping. The data are then converted into NumPy arrays for further calculation, and users can apply filters.

The tool has been available for a few months, and "it has already been used by teams around the world," Dr Aerts reported at the Radiological Society of North America 2017 Annual Meeting in Chicago. "We're pretty proud of this project."

Because researchers can contribute to the tool, the community can work together.

State-of-the-Art Feature Extraction Algorithms

"It's open source. The more attention it gets, the more valuable it becomes. It's not ours; it's everybody's," he explained.

In a recent report, Dr Aerts and his colleagues describe the platform and demonstrate its application in a case study designed to distinguish between benign and malignant lung lesion nodules (Cancer Res. 2017;77:e104-e107).

The researchers extracted 1120 radiomics features — 14 shape features, 19 first-order intensity statistics features, 60 texture features, 395 Laplacian of Gaussian features, and 632 wavelet features — from 429 lesions.

The image-based multivariate biomarker the team developed demonstrated a strong and significant ability to characterize lung nodules in the validation cohort (area under the curve [AUC], 0.79; Noether test, P = 4.12e−22).

Others have also reported success with the tool. Researchers from the Netherlands Cancer Institute compared machine and human analyses of MRI images to detect cancer in the rectum.

"They showed me results that blew my mind. The machine analysis results were as good as the human analysis," Dr Aerts reported.

We can now analyze things in a few weeks that would have once taken years.

Deep learning will eventually pass the threshold of clinical analysis. "This is going to be useful for diagnosis, detection, monitoring, heart disease, cancer, neurologic diseases," he said. "We can now analyze things in a few weeks that would have once taken years."

Speed is the big draw of radiomics. The ability to extract and analyze a large number of advanced quantitative features from medical images obtained with CT, PET, or MRI is staggering. The hope is that radiomics artificial intelligence technology — either engineered hard-coded algorithms or deep-learning methods — will be used to develop noninvasive imaging-based biomarkers.

It could mean a lot fewer biopsies and more precise treatment.

"With this resource, we aim to establish a reference standard for radiomic analyses, provide tested and maintained open-source platforms, and raise awareness among scientists of the potential for radiomics technologies," Dr Aerts said.

Added Value of Radiomics

Radiomics is an important component for the integration of multimodal data for diagnostic and prognostic applications, said Neema Jamshidi, MD, PhD, from the Ronald Reagan UCLA Medical Center in Los Angeles.

However, artificial intelligence should not be seen as a panacea for all technical challenges, he cautioned.

"Realistically, what we have right now is a number of studies showing a proof of principle for the utility of these approaches that milk more data out of the images," he explained. These quantitative measures are providing added value, "but you still have to actually drill down, understand, and figure out what the important things to measure are, and how can you measure them more efficiently."

Dr Jamshidi described two studies he was involved with. An analysis of women with breast cancer revealed an association between a radiogenomic biomarker — which included dynamic contrast-enhanced MRI and long noncoding RNA expression — and metastasis (Radiology. 2015;275:384-392).

And a phase 2 clinical trial of renal cell cancer showed that outcomes could be stratified with a radiogenomic risk score (Eur Radiol. 2016;26:2798-2807). The study "used a predictor based on expert-read imaging features, showing you could potentially be more efficient when using radiomic-driven traits," he reported.

But to actually demonstrate added value in clinical trials, there must be an objective assessment of how predictive the processes are and how much they decrease costs.

"Instead of looking at 10 or 20 features of an image, you can now look at 100," he pointed out. "But what does this actually mean?"

Dr Jamshidi said, "I want to see this progress and succeed, but we have to avoid biases and not go into presale mode."

Dr Aerts and Dr Jamshidi have have disclosed no relevant financial relationships.

Radiological Society of North America (RSNA) 2017 Annual Meeting. Abstract SPSH20B. Presented November 27, 2017.

Follow Medscape on Twitter @Medscape and Ingrid Hein @ingridhein


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.