AI Controversies | RB-82 Warning | Pigg-O-Stat Goes Viral

“Everyone who argues that ‘AI isn’t magic’ needs to have an infusion of childlike wonder, stat.”

Radiologist and PhD Candidate (and 2nd time Imaging Wire quote provider), Luke Oakden-Rayner, touting the qualities of the term “Artificial Intelligence” and AI’s magical ability to transform images into decisions.


Imaging Wire Sponsors

  • Carestream – Focused on delivering innovation that is life changing – for patients, customers, employees, communities and other stakeholders.
  • Focused Ultrasound Foundation – Accelerating the development and adoption of focused ultrasound.
  • Medmo – Helping underinsured Americans save on medical scans by connecting them to imaging providers with unfilled schedule time.
  • Nuance – AI and cloud-powered technology solutions to help radiologists stay focused, move quickly, and work smarter.
  • Pocus Systems – A new Point of Care Ultrasound startup, combining a team of POCUS veterans with next-generation technology to disrupt the industry.
  • Qure.ai – Making healthcare more accessible by applying deep learning to radiology imaging.

The Imaging Wire

Ten Controversial AI Opinions
Dr. Luke Oakden-Rayner released another excellent AI blog that has people in the radiology and artificial intelligence arenas talking. Oakden-Rayner targeted some largely agreed-upon beliefs and approaches, and although he warns readers to take his opinions “with salt for best results,” we’ve found his perspectives usually come well-seasoned. Here they are:

  1. Open data is not necessarily good. Companies lose their competitive advantage by making their data open and open data is a “terrible thing for generalizability.”
  2. “Normal vs abnormal is a terrible task to train a model for. The abnormal class is so broad and diverse that your data will never cover it well”
  3. “Artificial intelligence” is a great term. Despite many people’s efforts to create distance from the term “artificial intelligence,” it’s a great term because “it brings in interest and money to the field, and frankly what we do is magic so let’s just run with it.”
  4. Deep learning is pretty useless for EHR data. This is emerging as one of the more controversial of Luke’s 10 AI opinions, which argues that deep learning isn’t that great with unstructured data like EHR records, so we shouldn’t expect DL/EHR breakthroughs any time soon.
  5. End-user interpretability is overrated. Even if many in healthcare say they want interpretable AI, “most doctors will gleefully and immediately cede all related decision making to the AI” (if it works), and interpretability methods will (at best) provide a false sense of security.
  6. No medical advance is going to be achieved by a team who has designed a fancy new model for the task. If a team builds its own model for medical research instead of using an already-available AI model, they’re actually doing machine learning research because “the very process of building and tuning your own model means you will almost certainly overfit to your particular data.”
  7. Releasing public code is not particularly relevant in medical AI research. In order for public code to improve reproducibility, everyone else would have to have an “equally good (but different) dataset” to validate results.
  8. Vision is done and dusted. Oakden-Rayner is forecasting the end of computer vision, suggesting that computer vision model performance isn’t going to get a lot better going forward.
  9. Unsupervised learning isn’t clinically relevant. AI currently requires human supervision to add value, and even if unsupervised learning is improving, it will “always take a performance hit, which will always make it worse than human.” In other words, “we won’t be solving medicine with our huge stores of unlabelled data anytime soon.”
  10. Distrust any system with an AUC below 0.8, since this score “is roughly how well medical AI systems work when they overfit on non-pathological image features” and “these systems will mostly fail as clinical AI because they can’t generalize.”


RB-82 Warning
Following recent incidents where imaging staff used incorrect rubidium-82 generator solutions, leading to excess radiation exposure during PET scans, the FDA issued a warning and instituted new safety labeling requirements. The FDA now requires a Boxed Warning on rubidium 82 generators that reminds operators to use the correct solution to elute the generator (to extract Rubidium 82) and to carefully follow safety procedures.

The reported incidents used calcium-containing solutions such as lactated ringers to elute the system, instead of the required 0.9% Sodium Chloride Injection USP. In these cases, the calcium can interact with the generator’s radioactive strontium 82 and strontium 85 isotopes, which are then inadvertently injected into the patients along with the rubidium, potentially leading to a range of medical issues (suppressed bone marrow and immune system functions, radiation-induced cancers).

Some have been quick to point out that these types of mistakes should never happen (with or without warning labels), but the fact that they are happening suggests that these extra steps are indeed necessary.

Q1 2019 Medical Imaging Financials off to Mixed Start
The first round of medical imaging company financials from the January-March 2019 period hit the press, revealing solid imaging division results from GE, Phillips, and Siemens Healthineers, along with (perhaps seasonal) downturns from a number of major players.

  • CanonCanon struggled in Q1 as its revenue fell by 10% to ¥864.5 billion ($7.7b) and its net income sunk by 45.2% to ¥31.3 billion ($281m – still pretty good). Like nearly every other Canon business unit, Canon Medical Systems struggled in Q1 (revenue -6.5% to ¥109.4b/$984m; OP -36% to ¥6.7b/$60m), ending a streak of three straight positive quarters for the medical division.

  • GEGE beat expectations again in Q1 (revenue -2% to $27.3b, net earnings +125% to $0.9B, FCF improved by $0.5b to -$1.2b), while GE Healthcare remained strong and stable (revenue flat at $4.7B, profit +6% to $0.8b).

  • Hitachi – Hitachi didn’t release its Q4 financials, but its full 2018 fiscal year that ended in March brought a 1% revenue increase to ¥9.48 trillion ($85.3b) and 34.6% drop in net income to ¥321 billion ($2.89b – yes, Hitachi is that big). Hitachi’s healthcare business had ¥176 billion ($1.58b) in revenue and ¥4.3 billion ($38.7m) in operating income during the full fiscal year (healthcare comparison vs. 2017 unavailable).

  • HologicHologic’s fiscal Q2 brought a 3.7% revenue increase to $818.4 million, driven in part by a 7.1% increase in breast health revenue to $321.5 million and an 11.4% increase in Molecular Diagnostics revenue of $167.8 million, while its GAAP loss improved to a $272.6 million net loss (vs. -$681.4m).

  • MednaxMednax’s Q1 saw a 1% increase in revenue to $851.2 million, while posting a $242.9 million net loss (vs. +$63.4m in Q1 2018) and an EBITDA of $113.6 million (vs. $133m in Q1 2018).

  • SamsungSamsung struggled in Q1 as revenue fell by 13.5% to KRW 52.39 trillion ($45b) and operating profit plummeted 60% to KRW 6.23 trillion ($5.35b) due to challenges in its memory and display businesses. Samsung’s Consumer Electronics division (which somehow includes healthcare) saw revenue increase 3% to KRW 10 trillion ($8.58b) and operating profit rise slightly to KRW 54 billion ($46m).

  • Siemens Healthineers – A strong imaging performance (+7% to €2.137/$2.395b) drove Siemens Healthineers’ 5.8% Q2 revenue growth to €3.505b ($3.909) and a 24% net income increase to €381m ($427m). Siemens’ imaging business appears to have outperformed its big four rivals (and the overall market), due in part to strong performance in Europe.


Breast SWE
A team of researchers from France and Quebec found that shear wave elastography (SWE) ultrasound may be an effective method for differentiating benign from malignant microcalcifications that are first detected on mammograms.

The study looked at 74 patients with mammographically detected suspicious microcalcifications who underwent breast US and had the microcalcifications assessed with SWE. The SWE assessments revealed that the group’s 13 malignant calcifications had “significantly higher” stiffness than the 16 benign calcifications. SWE diagnosed malignancy with a 0.89 AUC, 100% specificity, 69% sensitivity, 80% negative predictive value, 100% positive predictive value, and 86% accuracy.

Pigg-O-Stat Goes Viral
The Pigg-O-Stat Pediatric Immobilizer, a nearly 60-year-old contraption used to keep children still during X-ray exams, went viral last week when many in the non-medical community caught their first glimpse of how kids look when they’re strapped into the Pigg-O-Stat (cute, but definitely not comfortable). Reactions to the Pigg-O-Stat swept across social media and online publications, perhaps making last week the awareness high point for pediatric radiology and radiation safety in recent memory (this is the reason we’re covering it). That said, it’s worth noting that the Pigg-O-Stat also went viral for the same reasons in 2015.


The Wire

  • The Hong Kong Polytechnic University (PolyU) developed a handheld 3D ultrasound imaging system that’s intended for mass scoliosis screening and is positioned as a possible X-ray alternative. The new 5kg “Scolioscan Air” is a scaled-down version of PolyU’s 150kg Scolioscan system that launched in 2016, combining a wireless ultrasound probe with an optical marker, a depth camera, and a laptop or tablet PC with dedicated 3D spinal deformity analysis software to detect and measure signs of scoliosis.

  • A single-center study from a NYU led team reveals that up to 37% of CT pulmonary angiography (CTPA) scans for potential pulmonary embolism (PE) were performed without the support of imaging order guidelines. The early 2016 study looked at 212 consecutive CTPA scans and found that between 53 (25%) and 79 (37%) studies were guideline-discordant (depending on the scoring system), with 46 (22%) studies discordant under all three scoring systems used in the study. However, this may not be as reckless as it may seem, as 18 (39%) of the discordant scans “had at least one patient-specific factor associated with increased risk for PE but not included in the risk stratification scores (eg, travel, thrombophilia).”

  • RefleXion Medical closed a $60 million loan to fund the FDA clearance and commercial launch of its biology-guided radiotherapy (BgRT) system. The novel system uses PET to determine tumor location and then BgRT responds to emissions produced by the PET’s FDG tracer, potentially treating several sites simultaneously with less toxicity.

  • Research from a Medical University of South Carolina-led team found that an automated algorithm can accurately predict major adverse cardiovascular events (MACE) by analyzing plaque characteristics in CCTA scans with more accuracy than using traditional clinical risk factors. The study looked at 45 patients with suspected coronary artery disease (16 experienced MACE in the the last year) who underwent CCTA, achieving 77% accuracy and a 0.94 AUC with the algorithm-alone and 87% accuracy and an AUC of 0.94 by combining the algorithm with clinical risk factors (vs. 63% and 0.587 AUC with just clinical factors). The researchers suggested that the algorithm’s ability to reduce variability and improve efficiency could help CCTA achieve a more mainstream clinical role.

  • British Columbia’s health minister revealed impressive results from its efforts to improve MRI accessibility, increasing the total number of MRI scans performed in the province in the last year by 23% (43,993 additional scans) and increasing scan volume in some areas by almost 87%. These increases were made possible by shifting 10 of the province’s 33 MRI machines to a 24×7 schedule, buying two privately owned MRI clinics in the Fraser Valley (east of Vancouver), and increasing its MRI budget by $11 million. The BC government will add another $5.25 million to its MRI budget next year, funding 15,000 additional MRI scans.

  • Research from a UCSF and Mayo Clinic team found that breast ultrasound’s high NPV makes it the best initial imaging option for palpable lumps after mastectomy. The study looked at 101 patients (118 cases, 75 with ultrasound and 43 with US+DM), identifying nine cancers with only ultrasound and three cancers using US+DM. Of the 104 palpable nonmalignant lumps, ultrasound had a 97% negative predictive value and a 27% positive predictive value 2 (PPV2).

  • Chinese researchers developed a deep learning model that might be able to confirm, locate, and measure chronic myocardial infarction using cardiac cine MRI without the use of gadolinium. The study used cardiac cine MRI scans from 212 patients (80% for training, 20% for testing), with the DL model achieving respective per-segment sensitivity and specificity of 89.8% and 99.1%, and a 0.94 AUC.


The Resource Wire

This is sponsored content.

  • This Qure.ai blog details the challenges it overcame with its qER Head CT algorithm, including developing the algorithm to support CT’s 3D images and high resolution as well as the validation hurdles it faced due to the low prevalence of abnormalities, the need to create a dataset enriched with positives, and steps required to support radiologist reading.

  • How much does a CT scan cost? According to Medmo, that depends. Scans made with the exact same device on the exact same body part could cost $225 at one facility and $2,500 at another. Medmo also provides some advice to make sure patients don’t pay too much for their scans, including using the Medmo Marketplace, where the average CT costs between $225 and $700.

Leave a Reply

Basic HTML is allowed. Your email address will not be published.

Subscribe to this comment feed via RSS

Get every issue of The Imaging Wire

Delivered right to your inbox. We will never sell or share your information.