AI-powered cancer diagnoses for resource limited settings

An affordable point-of-care system integrated with microholography and deep learning allows for rapid and accurate molecular diagnosis of aggressive lymphomas.

Like Comment

The development of affordable diagnostic systems for low resource settings and global healthcare represents a key interest of the translational team at the Center for Systems Biology at Massachusetts General Hospital (MGH). The team, led by Dr. Ralph Weissleder, has developed a variety of such systems for oncologic and infectious applications, among other groups of diseases1–5. Given my experience in nanoplasmonic and photonic sensors, I was naturally attracted to the medical opportunities that these new generation of devices could address. 

When I first joined the team, we leveraged smartphones to implement immuno-microholography for cervical cancer diagnostics6. While feasible, it soon became clear that this approach had its own limitations (e.g. increasingly expensive smartphones with rapid development cycles, limited control to proprietary hardware). We thus set out to design, develop and iteratively test novel standalone systems that incorporate the latest image sensors and algorithms. The challenge was to design such systems at reasonable costs (i.e. lower than that of a smartphone), with integrated computing power and with a software algorithm that could be used by unskilled healthcare workers. The reported work represents the culmination of several years of development and validates the technology in a clinical trial in patients with suspected lymphoma7. The current data show the superiority of our device over very expensive flow cytometry systems. In the future, we envision a broader application of this technology for other cancers, including those of the breast or HIV-associated malignancies which are prevalent in low and middle-income countries (LMICs). 

1. Stand-alone device
The basic principle behind immuno-microholography is immunostaining (e.g. with chromogens7 or beads6) and/or immunocapturing of cells on glass surfaces. The holographic diffraction patterns of ‘stained’ cells differs from that of unstained cells, thus allowing for molecular  diagnoses. Advantages over immunocytology lie in the larger field of view, higher throughput, and lower dependency on specialists (AI powered algorithms obviate full reliance on highly trained pathologists). In all, its automatic operation, high throughput and low cost render the technology ideally suited for cellular analyses in low resource settings. 

The new stand-alone device (Fig. 1) represents an integrated system capable of data acquisition, communication, analysis and display. The device outperforms the previous iPhone-based device in optical and computational performance at a lower cost. It also enables the use of disposable microfluidic cartridges to simplify assay procedures. The standalone device provides us more flexibility to customize setups and adds new functions. More importantly, it allows for running deep-learning algorithms for automated on-site analyses.

Figure 1. Stand-alone CEM device, one of the lowest-cost molecular diagnostic systems, consists of imaging, computing, communication and user-interface modules. The image was adapted from Ref. 7.

2. Deep-learning algorithm
Applying deep-learning algorithms for rapid and automated analysis was one of the key additions to the new system. Conventional microholography involves image reconstruction to retrieve microscopic images of objects. This is a computationally intensive process that requires high-end computing with graphic processing units8. Previously, we used cloud computing with wireless data transfer6. This approach, however, is suboptimal in resource limited settings as high-speed data connection is not ubiquitously available, can be expensive and can be a source of error (owing to dropped communications). We reasoned that deep-learning analysis on diffraction patterns could overcome such challenges. Our deep-learning algorithms based on convolutional neural networks showed that it could accurately detect lymphoma cells using a standalone device to obtain diagnostic results in a few minutes (Fig. 2).

Figure 2. A deep learning algorithm based on a convolutional neural network (CNN) identified B cells directly from the holograms. Holographic subimages of single cells were used for training the CNN. The images were adapted from Ref. 7.

After several technological refinements, we validated the system and assay with a real-world clinical study of 40 patients enrolled at MGH. In comparison to flow cytometry, our approach based on contrast-enhanced microholography (CEM) showed higher accuracy (95%) than flow cytometry (87%). Furthermore, based on size measurements, which are not readily identifiable with flow cytometry, we were able to identify aggressive subtypes with an accuracy of 86%.

Looking forward, with support from the National Cancer Institute, we will deploy the device and conduct a clinical trial in hundreds of patients with mass lesions in Botswana. The system will be able to decentralize patient triage in areas where diagnostic turnaround time can take weeks to months. We also hope to conduct other trials for different cancers (e.g. breast, head and neck, and others) prior to rolling out this technology in other LMICs. 

Importantly, the described work reflects the type of multi-disciplinary efforts needed to disrupt existing medical delivery models for improved patient care. As an engineer working with a group of forward-thinking investigators, I am motivated to continue pushing the envelope with the hope of delivering new technologies to help the millions of populations facing illness.   

Our paper: Im, H et al., Design and clinical validation of a point-of-care device for the diagnosis of lymphoma via contrast-enhanced microholography and machine learning. Nat. Biomed. Eng. (2018) 10.1038/s41551-018-0265-3.


1. Haun, J. B. et al. Micro-NMR for rapid molecular analysis of human tumor samples. Sci Transl Med 3, 71ra16 (2011).

2. Im, H. et al. Label-free detection and molecular profiling of exosomes with a nano-plasmonic sensor. Nat Biotechnol 32, 490-495 (2014).

3. Min, J. et al. Integrated Biosensor for Rapid and Point-of-Care Sepsis Diagnosis. ACS Nano 12, 3378-3384 (2018).

4. Park, K. S. et al. Rapid identification of health care-associated infections with an integrated fluorescence anisotropy system. Sci Adv 2, e1600300 (2016).

5. Shao, H. et al. Chip-based analysis of exosomal mRNA mediating drug resistance in glioblastoma. Nat Commun 6, 6999 (2015).

6. Im, H. et al. Digital diffraction analysis enables low-cost molecular diagnostics on a smartphone. Proc Natl Acad Sci U S A 112, 5613-5618 (2015).

7. Im, H. et al., Design and clinical validation of a point-of-care device for the diagnosis of lymphoma via contrast-enhanced microholography and machine learning. Nat. Biomed. Eng. (2018) 10.1038/s41551-018-0265-3.

8. Greenbaum, A. et al. Imaging without lenses: achievements and remaining challenges of wide-field on-chip microscopy. Nat Methods 9, 889-895 (2012).

Hyungsoon Im

Assistant Professor, Massachusetts General Hospital